Use query lib - #823
Use query lib#823abnegate wants to merge 771 commits into
Conversation
|
Note Reviews pausedIt looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the Use the following commands to manage reviews:
Use the checkboxes below for quick actions:
📝 WalkthroughWalkthroughLarge refactor: Query delegates to utopia-php/query BaseQuery; adapters, traits, hooks, and new domain types/enums (Attribute, Index, Relationship, Capability, etc.) migrate to value-object and capability models; CI, Docker, and composer updated to use a local query path and Paratest. Changes
Sequence Diagram(s)sequenceDiagram
participant Client
participant Pool
participant Adapter
participant WriteHooks as "Write Hooks\n(PermissionWrite/TenantWrite)"
participant Relationship as "RelationshipHandler"
Client->>Pool: createDocument(collection, Document)
Pool->>Adapter: createDocument(Document)
Note right of Adapter: decorate row via write hooks
Adapter->>WriteHooks: decorateRow(row, metadata)
WriteHooks-->>Adapter: decorated row
Adapter->>Adapter: persist row (DB)
Adapter->>Relationship: afterDocumentCreate(collection, document)
Relationship-->>Adapter: possibly update/populate related docs
Adapter-->>Pool: Document (created)
Pool-->>Client: Document (created)
Estimated code review effort🎯 5 (Critical) | ⏱️ ~120 minutes Possibly related PRs
Suggested reviewers
✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
There was a problem hiding this comment.
Actionable comments posted: 2
🧹 Nitpick comments (1)
src/Database/Query.php (1)
62-76: Align cursor helper docblocks with the actual parameter type.Lines 62 and 73 document
Document, but the signature ismixed $value(Lines 65 and 76). Please update phpdoc to match actual behavior.🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Query.php` around lines 62 - 76, The PHPDoc for Query::cursorAfter and Query::cursorBefore incorrectly types the parameter as Document while the method signatures accept mixed $value; update the `@param` annotations in both docblocks to "@param mixed $value" (and keep "@return Query" unchanged) so the docblocks match the actual signatures for the cursorAfter and cursorBefore methods on the Query class.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@composer.json`:
- Around line 62-66: Replace the SSH Git URL in the composer.json "repositories"
entry with its HTTPS equivalent to avoid CI SSH auth issues; locate the
"repositories" array and update the url value currently set to
"[email protected]:utopia-php/query.git" to use
"https://github.com/utopia-php/query.git" so installs succeed in non-interactive
CI environments.
In `@src/Database/Query.php`:
- Around line 167-183: The code is using $limit as a fallback for both $offset
and $cursor which is incorrect; update the assignments in the TYPE_OFFSET and
TYPE_CURSOR_AFTER/TYPE_CURSOR_BEFORE branches so they use the actual first value
or null instead of $limit (i.e. change $values[0] ?? $limit to $values[0] ??
null) while keeping the existing checks that ignore subsequent offsets/cursors
and preserving $cursorDirection assignment (references: TYPE_OFFSET,
TYPE_CURSOR_AFTER, TYPE_CURSOR_BEFORE, $offset, $cursor, $values, $limit,
$cursorDirection, Database::CURSOR_AFTER, Database::CURSOR_BEFORE).
---
Nitpick comments:
In `@src/Database/Query.php`:
- Around line 62-76: The PHPDoc for Query::cursorAfter and Query::cursorBefore
incorrectly types the parameter as Document while the method signatures accept
mixed $value; update the `@param` annotations in both docblocks to "@param mixed
$value" (and keep "@return Query" unchanged) so the docblocks match the actual
signatures for the cursorAfter and cursorBefore methods on the Query class.
ℹ️ Review info
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
⛔ Files ignored due to path filters (1)
composer.lockis excluded by!**/*.lock
📒 Files selected for processing (2)
composer.jsonsrc/Database/Query.php
There was a problem hiding this comment.
♻️ Duplicate comments (1)
src/Database/Query.php (1)
161-177:⚠️ Potential issue | 🟠 MajorFix fallback leakage from
limitinto offset/cursor state.Line 167 and Line 176 incorrectly use
$limitas fallback, which can corrupt grouped query output when offset/cursor values are missing.🔧 Proposed fix
case self::TYPE_OFFSET: @@ - $offset = $values[0] ?? $limit; + $offset = $values[0] ?? $offset; break; @@ case self::TYPE_CURSOR_AFTER: case self::TYPE_CURSOR_BEFORE: @@ - $cursor = $values[0] ?? $limit; + $cursor = $values[0] ?? $cursor; $cursorDirection = $method === self::TYPE_CURSOR_AFTER ? Database::CURSOR_AFTER : Database::CURSOR_BEFORE; break;🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Query.php` around lines 161 - 177, In Query:: (switch handling TYPE_OFFSET / TYPE_CURSOR_AFTER / TYPE_CURSOR_BEFORE) remove the incorrect fallback to $limit so $offset and $cursor aren't polluted: instead of assigning $values[0] ?? $limit, only assign when an explicit value exists (e.g. check isset($values[0]) or array_key_exists) and otherwise leave $offset/$cursor as null; keep the first-occurrence guards intact and preserve setting $cursorDirection when handling TYPE_CURSOR_AFTER vs TYPE_CURSOR_BEFORE (Database::CURSOR_AFTER / Database::CURSOR_BEFORE).
🧹 Nitpick comments (1)
src/Database/Query.php (1)
53-73: Align cursor helper PHPDoc with the actualmixedparameter type.Both docblocks still document
@param Document $value, but the signatures now acceptmixed.📝 Suggested doc fix
- * `@param` Document $value + * `@param` mixed $value @@ - * `@param` Document $value + * `@param` mixed $value🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Query.php` around lines 53 - 73, Update the PHPDoc for the Query helper methods to match the actual parameter type: in the Query class update the docblocks for cursorAfter and cursorBefore so the `@param` annotation is "@param mixed $value" (instead of "@param Document $value") and keep the `@return` annotation as "Query"; ensure these docblocks sit immediately above the corresponding methods cursorAfter and cursorBefore to maintain accurate IDE/typehinting.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Duplicate comments:
In `@src/Database/Query.php`:
- Around line 161-177: In Query:: (switch handling TYPE_OFFSET /
TYPE_CURSOR_AFTER / TYPE_CURSOR_BEFORE) remove the incorrect fallback to $limit
so $offset and $cursor aren't polluted: instead of assigning $values[0] ??
$limit, only assign when an explicit value exists (e.g. check isset($values[0])
or array_key_exists) and otherwise leave $offset/$cursor as null; keep the
first-occurrence guards intact and preserve setting $cursorDirection when
handling TYPE_CURSOR_AFTER vs TYPE_CURSOR_BEFORE (Database::CURSOR_AFTER /
Database::CURSOR_BEFORE).
---
Nitpick comments:
In `@src/Database/Query.php`:
- Around line 53-73: Update the PHPDoc for the Query helper methods to match the
actual parameter type: in the Query class update the docblocks for cursorAfter
and cursorBefore so the `@param` annotation is "@param mixed $value" (instead of
"@param Document $value") and keep the `@return` annotation as "Query"; ensure
these docblocks sit immediately above the corresponding methods cursorAfter and
cursorBefore to maintain accurate IDE/typehinting.
There was a problem hiding this comment.
Actionable comments posted: 1
♻️ Duplicate comments (1)
src/Database/Query.php (1)
159-169:⚠️ Potential issue | 🟠 MajorFix incorrect fallback variable usage for offset and cursor.
Lines 159 and 168 incorrectly fall back to
$limitinstead of the appropriate variable. This can silently corrupt the grouped query state.
- Line 159:
$offset = $values[0] ?? $limitshould use$offset(ornull)- Line 168:
$cursor = $values[0] ?? $limitshould use$cursor(ornull)🔧 Proposed fix
case self::TYPE_OFFSET: // Keep the 1st offset encountered and ignore the rest if ($offset !== null) { break; } - $offset = $values[0] ?? $limit; + $offset = $values[0] ?? null; break; case self::TYPE_CURSOR_AFTER: case self::TYPE_CURSOR_BEFORE: // Keep the 1st cursor encountered and ignore the rest if ($cursor !== null) { break; } - $cursor = $values[0] ?? $limit; + $cursor = $values[0] ?? null; $cursorDirection = $method === self::TYPE_CURSOR_AFTER ? Database::CURSOR_AFTER : Database::CURSOR_BEFORE; break;🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Query.php` around lines 159 - 169, The switch handling in Query.php incorrectly falls back to $limit when assigning $offset and $cursor, corrupting grouped query state; update the assignments in the cases that set $offset and $cursor (the branches that assign $offset = $values[0] ?? $limit and $cursor = $values[0] ?? $limit) to use a proper null fallback instead (e.g. $offset = $values[0] ?? null and $cursor = $values[0] ?? null), keeping the surrounding logic for cursorDirection (Database::CURSOR_AFTER / Database::CURSOR_BEFORE) intact.
🧹 Nitpick comments (1)
src/Database/Query.php (1)
54-65: Consider tightening parameter type hint.The PHPDoc indicates
@param Document $value, but the signature usesmixed. Consider usingDocumentas the type hint for better type safety, or update the PHPDoc to reflect the actual accepted types.🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Query.php` around lines 54 - 65, The PHPDoc for cursorBefore indicates the parameter is a Document but the method signature (and cursorAfter) use mixed; update the signatures to use Document instead of mixed for cursorBefore and cursorAfter (i.e., change the parameter type from mixed to Document) to match the PHPDoc and improve type safety, or alternatively update the PHPDoc to reflect mixed if these methods truly accept other types—adjust the declarations in the Query class (cursorBefore, cursorAfter) so the docblock and method signatures are consistent.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@src/Database/Query.php`:
- Around line 42-49: The method parseQuery currently declares a return type of
static but calls parent::parseQuery() which returns BaseQuery; change the method
signature to return BaseQuery instead of static and keep the try/catch that
wraps BaseQueryException into QueryException (i.e., update the return type on
parseQuery to BaseQuery to match parent::parseQuery and ensure the thrown
QueryException still wraps the original BaseQueryException).
---
Duplicate comments:
In `@src/Database/Query.php`:
- Around line 159-169: The switch handling in Query.php incorrectly falls back
to $limit when assigning $offset and $cursor, corrupting grouped query state;
update the assignments in the cases that set $offset and $cursor (the branches
that assign $offset = $values[0] ?? $limit and $cursor = $values[0] ?? $limit)
to use a proper null fallback instead (e.g. $offset = $values[0] ?? null and
$cursor = $values[0] ?? null), keeping the surrounding logic for cursorDirection
(Database::CURSOR_AFTER / Database::CURSOR_BEFORE) intact.
---
Nitpick comments:
In `@src/Database/Query.php`:
- Around line 54-65: The PHPDoc for cursorBefore indicates the parameter is a
Document but the method signature (and cursorAfter) use mixed; update the
signatures to use Document instead of mixed for cursorBefore and cursorAfter
(i.e., change the parameter type from mixed to Document) to match the PHPDoc and
improve type safety, or alternatively update the PHPDoc to reflect mixed if
these methods truly accept other types—adjust the declarations in the Query
class (cursorBefore, cursorAfter) so the docblock and method signatures are
consistent.
There was a problem hiding this comment.
Actionable comments posted: 3
Note
Due to the large number of review comments, Critical severity comments were prioritized as inline comments.
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (6)
src/Database/Document.php (1)
232-243:⚠️ Potential issue | 🟡 MinorUpdate docblock to match the new parameter type.
The docblock still documents
@param string $typebut the actual parameter type is nowSetType. This could confuse IDE autocompletion and static analysis tools.📝 Proposed fix
/** * Set Attribute. * * Method for setting a specific field attribute * * `@param` string $key * `@param` mixed $value - * `@param` string $type + * `@param` SetType $type * * `@return` static */ public function setAttribute(string $key, mixed $value, SetType $type = SetType::Assign): static🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Document.php` around lines 232 - 243, Update the docblock for the setAttribute method to match the new parameter type: replace the incorrect "@param string $type" with "@param SetType $type" (or fully qualified \Namespace\SetType if necessary) so IDEs and static analyzers reflect the actual signature of setAttribute(string $key, mixed $value, SetType $type = SetType::Assign) and keep the rest of the description intact.src/Database/Mirror.php (1)
398-422:⚠️ Potential issue | 🟠 MajorUse the filtered key when mirroring attribute renames.
beforeUpdateAttribute()can rewrite the attribute document, but the destination update still uses the caller’s raw$newKey. Any filter that renames destination attributes will drift on update.🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Mirror.php` around lines 398 - 422, The destination update call is still using the original caller $newKey even though writeFilters->beforeUpdateAttribute may have rewritten the attribute document (including renaming the key); update the call in Mirror:: where updateAttribute is invoked to use the filtered key from the mutated $document (e.g., $document->getAttribute('key')) instead of $newKey, falling back to $newKey if the document does not contain a key, so attribute renames applied by beforeUpdateAttribute are honored; reference writeFilters, beforeUpdateAttribute, $document, updateAttribute and $newKey when making the change.src/Database/Adapter/Pool.php (1)
101-104:⚠️ Potential issue | 🟠 MajorPersist the timeout on the pool wrapper.
This override forwards the current call but never updates
$this->timeout. Laterdelegate()calls therefore won’t reapply the timeout to newly borrowed adapters, so the setting is lost across pool checkouts.⏱️ Proposed fix
public function setTimeout(int $milliseconds, string $event = Database::EVENT_ALL): void { + parent::setTimeout($milliseconds, $event); $this->delegate(__FUNCTION__, \func_get_args()); }🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Adapter/Pool.php` around lines 101 - 104, The setTimeout override in Pool::setTimeout(int $milliseconds, string $event = Database::EVENT_ALL) forwards the call to delegate() but fails to persist the value to the pool wrapper, so $this->timeout is never updated and newly borrowed adapters don't get the timeout; update the Pool::setTimeout implementation to assign the passed $milliseconds (and $event if you store per-event) to the pool's internal property ($this->timeout or a per-event map) before or after calling $this->delegate(__FUNCTION__, func_get_args()), ensuring future delegate() calls reapply the stored timeout to adapters checked out from the pool.src/Database/Query.php (1)
240-257:⚠️ Potential issue | 🟠 MajorCursor queries stop working after
toArray()/groupForDatabase()round-trips.
Documents::find()expects$cursorto be aDocumentso it can read order keys and$collection.toArray()collapses cursor documents to$id, andgroupForDatabase()returns that scalar unchanged, so any serialized cursor query will fail when pagination code callsgetAttribute()orgetCollection()on it.🔧 Proposed fix
- if ($value instanceof Document && in_array($this->method, [Method::CursorAfter, Method::CursorBefore])) { - $value = $value->getId(); + if ($value instanceof Document && \in_array($this->method, [Method::CursorAfter, Method::CursorBefore], true)) { + $value = $value->getArrayCopy(); } $array['values'][] = $value; } } @@ + $cursor = $grouped->cursor; + if (\is_array($cursor)) { + $cursor = new Document($cursor); + } + return [ 'filters' => $filters, 'selections' => $selections, 'limit' => $grouped->limit, 'offset' => $grouped->offset, 'orderAttributes' => $grouped->orderAttributes, 'orderTypes' => $orderTypes, - 'cursor' => $grouped->cursor, + 'cursor' => $cursor, 'cursorDirection' => $cursorDirection, ];Also applies to: 280-317
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Query.php` around lines 240 - 257, The cursor-serialization code in Query::toArray() currently replaces Document instances with their id for cursor methods (Method::CursorAfter, Method::CursorBefore), which breaks Documents::find() that expects full Document objects after groupForDatabase() round-trips; change the branch so that when $value instanceof Document you serialize it with $value->toArray() (not ->getId()) so the document structure is preserved and can be rehydrated on read (also apply the same fix in the analogous block around lines 280-317), making sure groupForDatabase()/unserialize logic continues to round-trip these document arrays back into Document objects so getAttribute() and getCollection() still work.src/Database/Adapter/Mongo.php (2)
1575-1592:⚠️ Potential issue | 🟠 MajorKeep replacement-style updates for schemaless collections.
This now always wraps the payload in
$set. When defined attributes are disabled, missing keys are supposed to disappear on update; with$setthey persist forever because the old document is never replaced.Suggested fix
$options = $this->getTransactionOptions(); - $updateQuery = [ - '$set' => $record, - ]; + $updateQuery = $this->supportForAttributes + ? ['$set' => $record] + : $record; $this->client->update($name, $filters, $updateQuery, $options);Based on learnings: In src/Database/Adapter/Mongo.php, when getSupportForAttributes() returns false (schemaless mode), the updateDocument method intentionally uses a raw document without $set operator for replacement-style updates, as confirmed by the repository maintainer ArnabChatterjee20k.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Adapter/Mongo.php` around lines 1575 - 1592, The updateDocument method currently always wraps the payload in a $set which prevents replacement-style updates for schemaless collections; change updateDocument so that after preparing $record (and unsetting '_id') it checks $this->getSupportForAttributes(): if true keep $updateQuery = ['$set' => $record], but if false use $updateQuery = $record (a replacement document) before calling $this->client->update($name, $filters, $updateQuery, $options) so missing keys are removed for schemaless collections; refer to updateDocument and getSupportForAttributes() to locate where to branch.
1542-1553:⚠️ Potential issue | 🟠 MajorScope the post-insert readback.
In shared-table mode
_uidis only unique within_tenant. This follow-upfind()reads back by_uidalone, so a custom ID collision can hydrate the newly created document from another tenant's row instead of the one that was just inserted.🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Adapter/Mongo.php` around lines 1542 - 1553, The post-insert readback in insertDocument uses a filter of ['_uid' => $document['_uid']], which can collide across tenants; update the filter used in the inner client->find call to scope by both '_uid' and '_tenant' (e.g. include '_tenant' => $document['_tenant'] when present) so the find() returns the document from the same tenant that was just inserted, adjusting any variable names (filters) used around client->find accordingly.
🟠 Major comments (19)
composer.json-64-72 (1)
64-72:⚠️ Potential issue | 🟠 MajorPath repository will fail in CI environments.
The path repository pointing to
../queryis only valid for local development where the siblingquerydirectory exists. In CI environments (GitHub Actions, etc.) and for other contributors cloning this repo, this path won't exist andcomposer installwill fail.For CI compatibility, either:
- Publish
utopia-php/queryto Packagist before merging- Use the official VCS repository URL as a fallback
🔧 Proposed fix using VCS repository
"repositories": [ { "type": "path", "url": "../query", "options": { "symlink": true } - } + }, + { + "type": "vcs", + "url": "https://github.com/utopia-php/query.git" + } ],Note: Composer will prefer the path repository if available, falling back to VCS otherwise.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@composer.json` around lines 64 - 72, The current "repositories" entry uses a path repo ("type": "path", "url": "../query") which will break CI because the sibling ../query won't exist; update composer.json to remove or guard the path-only repository and add a fallback VCS repository entry pointing to the public repo URL for utopia-php/query (or publish to Packagist and use its package name), so Composer can use the VCS source in CI while still preferring the local path when present; modify the "repositories" array to include the official VCS URL (or switch to Packagist) alongside or instead of the existing path entry.src/Database/Hook/MongoPermissionFilter.php-26-27 (1)
26-27:⚠️ Potential issue | 🟠 MajorEscape regex metacharacters in role names to prevent regex injection.
Role names are directly interpolated into the regex pattern without escaping. If a role contains regex metacharacters (e.g.,
(,),|,.,*), it could break the regex or enable ReDoS attacks.🛡️ Proposed fix to escape role names
- $roles = \implode('|', $this->authorization->getRoles()); + $roles = \implode('|', \array_map( + fn($role) => \preg_quote($role, '/'), + $this->authorization->getRoles() + )); $filters['_permissions']['$in'] = [new Regex("{$forPermission}\\(\".*(?:{$roles}).*\"\\)", 'i')];🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Hook/MongoPermissionFilter.php` around lines 26 - 27, The code builds a Regex from role names without escaping them, risking regex injection; update the logic that creates $roles (from $this->authorization->getRoles()) to escape each role with a regex-escaping function (e.g., preg_quote) before joining them, then use the escaped string when constructing the Regex used in the $filters['_permissions']['$in'] assignment (the Regex("{$forPermission}\\(\".*(?:{$roles}).*\"\\)", 'i') instantiation) so special characters in role names are neutralized.src/Database/Traits/Attributes.php-998-1015 (1)
998-1015:⚠️ Potential issue | 🟠 MajorRollback the previous default when metadata persistence fails.
If the adapter update succeeds and
updateMetadata()then fails, the rollback model recreates the old column without its previous default. The physical schema can remain mutated while metadata rolls back.🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Traits/Attributes.php` around lines 998 - 1015, The rollback Attribute is missing the original default value so if the adapter update succeeds but updateMetadata() fails the physical column loses its default; modify the rollback model creation (the Attribute instance assigned to $rollbackAttrModel) to include the previous default (e.g., $originalDefault) and ensure whatever variable holds the prior default is captured and passed into the Attribute constructor, then keep the rollback closure that calls $this->adapter->updateAttribute($collection, $rollbackAttrModel, $originalKey) unchanged so the adapter will restore the column including its original default when metadata persistence fails.src/Database/Traits/Relationships.php-893-902 (1)
893-902:⚠️ Potential issue | 🟠 MajorPreserve the original side during rollback.
This recreation hard-codes
RelationSide::Parent. If a child-side delete fails after the physical schema change, rollback will restore the relationship on the wrong side.🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Traits/Relationships.php` around lines 893 - 902, The rollback recreates the Relationship with a hard-coded RelationSide::Parent which can restore the relation on the wrong side; change the code that builds $recreateRelModel in the rollback to use the original relationship side (e.g., the existing side value from the original Relationship instance or the local $side variable) instead of RelationSide::Parent so the recreated Relationship preserves the original side; update the constructor call that sets side to use that original side source (RelationSide value obtained from the original model) when constructing $recreateRelModel.src/Database/Traits/Attributes.php-281-284 (1)
281-284:⚠️ Potential issue | 🟠 MajorRollback should only delete attributes created in this batch.
$attributeDocumentsalso contains schema-only orphans that already existed before this call. If one new attribute is created andupdateMetadata()fails,cleanupAttributes()will delete those pre-existing columns too.Also applies to: 317-323
src/Database/Traits/Attributes.php-546-559 (1)
546-559:⚠️ Potential issue | 🟠 MajorInvalidate both collection and metadata caches after these mutations.
createAttribute(),updateAttribute(), anddeleteAttribute()purge cached collection state and the metadata document.updateAttributeMeta()does neither, andrenameAttribute()still leaves the metadata document cache warm, so callers can read stale attribute definitions after a successful write.Also applies to: 1284-1292
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Traits/Attributes.php` around lines 546 - 559, After performing attribute-mutating operations (specifically in updateAttributeMeta() and renameAttribute()), ensure you clear both the collection cache and the metadata document cache just like createAttribute()/updateAttribute()/deleteAttribute() do: call the same cache invalidation logic used elsewhere after the $this->updateMetadata(...) call (and before returning), and also ensure any cache keys for the metadata document are purged so callers don't read stale attribute definitions; reference updateAttributeMeta(), renameAttribute(), and the existing $this->updateMetadata(...) / self::EVENT_ATTRIBUTE_UPDATE flow to locate the correct spot to add the invalidation.src/Database/Adapter.php-685-697 (1)
685-697:⚠️ Potential issue | 🟠 MajorDon’t report unsupported relationship operations as success.
Traits\Relationshipsuses these return values to decide whether to update metadata. Returningtruehere lets an adapter without a real implementation claim success while leaving the physical schema unchanged.🧱 Proposed fix
public function createRelationship(Relationship $relationship): bool { - return true; + return false; } public function updateRelationship(Relationship $relationship, ?string $newKey = null, ?string $newTwoWayKey = null): bool { - return true; + return false; } public function deleteRelationship(Relationship $relationship): bool { - return true; + return false; }🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Adapter.php` around lines 685 - 697, The Adapter currently reports success for unsupported relationship operations; update the methods createRelationship(Relationship $relationship), updateRelationship(Relationship $relationship, ?string $newKey = null, ?string $newTwoWayKey = null) and deleteRelationship(Relationship $relationship) to not claim success — return false (or throw a clear UnsupportedOperationException) instead of returning true so Traits\Relationships won't assume metadata was updated; ensure the change is applied to those method bodies and keep the signatures unchanged.src/Database/Traits/Relationships.php-518-579 (1)
518-579:⚠️ Potential issue | 🟠 MajorUpdate both relationship metadata documents atomically.
These
updateAttributeMeta()calls commit independently. If the first collection update succeeds and the second one fails, the catch only reverts the adapter rename; it never restores the already-persisted metadata change on the first collection.🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Traits/Relationships.php` around lines 518 - 579, The metadata updates via updateAttributeMeta on the primary collection, the related collection (oldTwoWayKey), and the junction (when ManyToMany) must be performed atomically; currently if one update fails the others remain persisted. Fix by either wrapping all metadata updates and the adapter rename in a single DB transaction (begin/commit/rollback) so any failure rolls back all changes, or if transactions are not available, capture the existing attribute state for collection->getId()/$id, relatedCollection->getId()/oldTwoWayKey, and junction entries (if RelationType::ManyToMany) before applying updates and, inside the catch, call updateAttributeMeta to restore those saved states (and ensure purgeCachedCollection is similarly reverted). Update logic references: updateAttributeMeta, getJunctionCollection, withRetries->purgeCachedCollection, and the adapter->updateRelationship/Relationship model so both metadata and adapter rename are consistent.src/Database/Traits/Relationships.php-143-157 (1)
143-157:⚠️ Potential issue | 🟠 MajorCheck the related collection for
twoWayKeycollisions before creating the relationship.The duplicate scan only inspects the source collection.
checkAttribute()only enforces limits/width, so an existing attribute on the related collection with the same$twoWayKeyis not rejected before partial relationship creation starts.Also applies to: 191-192
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Traits/Relationships.php` around lines 143 - 157, The duplicate-check currently inspects only the source collection attributes (loop over $attributes) and misses collisions on the related collection; update the logic in the Relationships trait where you check for existing relationship attributes (the foreach that compares ColumnType::Relationship->value, twoWayKey and relatedCollection id) to also fetch and iterate the related collection's attributes (use $relatedCollection->getAttribute('attributes', [])) and compare their options['twoWayKey'] case-insensitively to $twoWayKey and their relatedCollection id to the source collection id, throwing DuplicateException('Related attribute already exists') on conflict; apply the same symmetric check to the other similar block referenced by checkAttribute() (the block around the other relationship-creation path) so both creation flows validate twoWayKey collisions on the opposite collection before proceeding.src/Database/Traits/Attributes.php-572-576 (1)
572-576:⚠️ Potential issue | 🟠 MajorKeep the metadata-only helpers behind the same invariants as
updateAttribute().
updateAttributeRequired()can mark an attribute required while leaving its old default in place,updateAttributeFilters()can remove mandatory filters likedatetime, andupdateAttributeDefault()never checks vector length against the configured dimension count. These helpers can persist attribute metadata that the full update path would reject.Also applies to: 627-631, 644-653
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Traits/Attributes.php` around lines 572 - 576, The three metadata-only helpers (updateAttributeRequired, updateAttributeFilters, updateAttributeDefault) bypass the same validation/invariants enforced by updateAttribute and can persist invalid metadata; change each helper to call the full update path or reuse its validation logic by delegating to updateAttribute (or to the common validation function used by updateAttribute) instead of directly mutating via updateAttributeMeta: ensure updateAttributeRequired validates compatibility between required flag and the existing default, updateAttributeFilters preserves/enforces mandatory filters (e.g., datetime) and validates any removed filters against attribute type, and updateAttributeDefault enforces vector length vs configured dimension count and any type-specific constraints before persisting. Ensure you reference and reuse updateAttribute or its shared validators so all invariants are consistently applied.src/Database/Hook/PermissionWrite.php-131-136 (1)
131-136:⚠️ Potential issue | 🟠 MajorRoute permission reads/deletes through
WriteContext::execute.The insert path uses the context executor, but these branches call
execute()directly and some of the delete helpers ignore the boolean result entirely. A failed permission mutation can therefore leave<collection>_permsout of sync with the document write.Also applies to: 187-192, 208-215, 227-231, 262-266
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Hook/PermissionWrite.php` around lines 131 - 136, The permission delete branches in PermissionWrite (use of ($context->newBuilder), $removeBuilder->delete(), and the subsequent ($context->executeResult) returning $deleteStmt) call $deleteStmt->execute() directly and ignore the WriteContext executor and result; change these to run deletes through the WriteContext executor (use WriteContext::execute on the prepared statement returned by ($context->executeResult) with Database::EVENT_PERMISSIONS_DELETE), check the boolean return value and propagate/handle failures (throw or abort the parent write) so permission mutations cannot silently fail; apply the same fix pattern to the other affected blocks that build/delete perms (the sections analogous to lines 187-192, 208-215, 227-231, 262-266).src/Database/Traits/Relationships.php-493-515 (1)
493-515:⚠️ Potential issue | 🟠 MajorDon’t treat the unchanged source key as proof that the rename already happened.
When only
$newTwoWayKeychanges,$actualNewKeystill points at the current source column. The orphan-recovery fallback will then mark$adapterUpdated = trueas soon as it sees that existing column, even if the related/junction rename never happened.🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Traits/Relationships.php` around lines 493 - 515, The recovery logic in Relationships.php incorrectly treats finding $actualNewKey in schema as proof the rename completed; instead, when catching the Throwable you must verify the rename truly occurred by checking that the new column exists AND the original source column was removed (or that the two-way/junction column change also exists) — use getSchemaAttributes(), $this->adapter->filter($actualNewKey) and also compare against the filtered original/source key (e.g., filteredOldKey) and/or the filtered $newTwoWayKey to ensure the found column isn't simply the unchanged source; only set $adapterUpdated = true if the new name exists AND the old name does not (or the corresponding two-way column was updated), otherwise rethrow the DatabaseException (preserving $e as previous).src/Database/Traits/Documents.php-970-975 (1)
970-975:⚠️ Potential issue | 🟠 MajorUse the hook's returned document in bulk updates.
updateDocument()reassignsafterDocumentUpdate(), butupdateDocuments()ignores the returned value. Any relationship hook that rewrites the payload will be skipped on this path.🔧 Proposed fix
$hook = $this->relationshipHook; if ($hook?->isEnabled()) { - $this->silent(fn () => $hook->afterDocumentUpdate($collection, $document, $new)); + $new = $this->silent(fn () => $hook->afterDocumentUpdate($collection, $document, $new)); } $document = $new;🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Traits/Documents.php` around lines 970 - 975, The bulk-update path is ignoring the relationship hook's return value: in the block using $this->relationshipHook and calling $hook->afterDocumentUpdate($collection, $document, $new), capture and use the hook's returned document (replace $new with the hook return when non-null) before assigning to $document so any payload rewrites by afterDocumentUpdate are preserved; update the logic around relationshipHook/isEnabled and the assignment to $document to use the hook's return value (from afterDocumentUpdate) instead of discarding it.src/Database/Traits/Documents.php-401-409 (1)
401-409:⚠️ Potential issue | 🟠 MajorRun
castingAfter()on single-document creates even when hooks are off.This path only normalizes adapter output inside the relationship-hook branch.
MariaDB::createDocument()returns the pre-cast document instance, so a single create can return DB-shaped values whilecreateDocuments()returns normalized ones.🔧 Proposed fix
$hook = $this->relationshipHook; if ($hook !== null && !$hook->isInBatchPopulation() && $hook->isEnabled()) { $fetchDepth = $hook->getWriteStackCount(); $documents = $this->silent(fn () => $hook->populateDocuments([$document], $collection, $fetchDepth)); - $document = $this->adapter->castingAfter($collection, $documents[0]); + $document = $documents[0]; } + $document = $this->adapter->castingAfter($collection, $document); $document = $this->casting($collection, $document); $document = $this->decode($collection, $document);🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Traits/Documents.php` around lines 401 - 409, The current code only runs $this->adapter->castingAfter(...) when $this->relationshipHook is present and enabled, causing single-document creates to return DB-shaped values; move or duplicate the adapter post-processing so castingAfter is applied for single-document flows regardless of the relationship hook state: ensure that after the potential hook population block you call $this->adapter->castingAfter($collection, $document) (using the same normalized $documents[0] result when hook ran) before running $this->casting(...) and $this->decode(...), so both createDocument() and createDocuments() return consistently normalized documents; use the existing symbols relationshipHook, populateDocuments, castingAfter, casting, decode and silent to implement this change.src/Database/Traits/Documents.php-274-301 (1)
274-301:⚠️ Potential issue | 🟠 MajorPreserve relationship roots when pruning selected fields.
$attributesToKeeponly records the literal selector. A query likeselect(['author.name'])will therefore remove the populatedauthorattribute itself, and aliased projections are dropped for the same reason. Please retain the root segment of dotted paths, plus any projection alias, before removing attributes.Based on learnings: Repo utopia-php/database: Relationship selects are always evaluated in the main alias context (no per-collection aliasing). In Utopia\Database\Database::applySelectFiltersToDocuments, do not rely on Query::getAlias() for relationships; instead, preserve the root of dotted paths and any projection alias (Query::getAs()) when filtering attributes.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Traits/Documents.php` around lines 274 - 301, The pruning logic in applySelectFiltersToDocuments builds $attributesToKeep from selectQueries by using literal selectors only, which drops relationship root attributes and aliased projections (e.g., select(['author.name']) or projections with Query::getAs()). Update the loop that populates $attributesToKeep (the foreach over $selectQueries and getValues()) to also: for each selector string preserve its root segment (the part before the first dot) and, if Query::getAs() is present for that query, preserve the alias name as well; do not rely on Query::getAlias() for relationship selects. Ensure these root keys and projection aliases are added to $attributesToKeep before the wildcard/internal-key checks so removeAttribute only strips truly unselected fields.src/Database/Traits/Documents.php-1174-1215 (1)
1174-1215:⚠️ Potential issue | 🟠 MajorDon't drop brand-new empty documents as a no-op.
When
$oldis empty and the incoming document has no user fields/operators,$hasChangesstays false and the item is removed from the batch.upsertDocument()then falls back togetDocument()and returns an empty document instead of creating a valid empty record.🔧 Proposed fix
- if (!$hasChanges) { + if (!$hasChanges && !$old->isEmpty()) { // If not updating a single attribute and the document is the same as the old one, skip it unset($documents[$key]); continue; }🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Traits/Documents.php` around lines 1174 - 1215, The current change-detection logic in the Documents trait can wrongly treat brand-new empty incoming docs as unchanged and drop them from $documents[$key], because when $old is empty and regularUpdatesUserOnly and $operators are empty $hasChanges remains false; update the logic in the block that computes $hasChanges (the checks that reference $operators, $attribute, $skipPermissionsUpdate, $old, regularUpdatesUserOnly, self::INTERNAL_ATTRIBUTES) to explicitly mark a document as changed when $old is null/empty (new record) even if no user fields/operators are present so upsertDocument() will create the empty record instead of falling back to getDocument(); ensure this new case still respects internal attribute filtering and the subsequent unset($documents[$key]) skip only applies to truly identical existing documents.src/Database/Adapter/MariaDB.php-603-609 (1)
603-609:⚠️ Potential issue | 🟠 MajorUse side-aware junction naming when renaming many-to-many columns.
deleteRelationship()computes the junction table name differently for parent vs child sides, butupdateRelationship()always uses<collection>_<related>. Child-side renames will target the wrong junction table.🔧 Proposed fix
- $junctionName = '_' . $collection->getSequence() . '_' . $relatedCollection->getSequence(); + $junctionName = $side === RelationSide::Parent + ? '_' . $collection->getSequence() . '_' . $relatedCollection->getSequence() + : '_' . $relatedCollection->getSequence() . '_' . $collection->getSequence();🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Adapter/MariaDB.php` around lines 603 - 609, updateRelationship() is using a fixed junction name ('_' . $collection->getSequence() . '_' . $relatedCollection->getSequence()) which ignores the relationship side and causes child-side renames to target the wrong junction table; change the junction name computation in updateRelationship() to mirror the logic used in deleteRelationship() (i.e., compute $junctionName based on the $side or whichever variable deleteRelationship() uses to decide order of $collection->getSequence() and $relatedCollection->getSequence()), then use that side-aware $junctionName when calling $renameCol for $key/$newKey and $twoWay/$newTwoWayKey so renames operate on the correct many-to-many table.src/Database/Adapter/MariaDB.php-1211-1228 (1)
1211-1228:⚠️ Potential issue | 🟠 MajorAdd handlers for
TYPE_COVERS,TYPE_NOT_COVERS,TYPE_SPATIAL_EQUALS, andTYPE_NOT_SPATIAL_EQUALSin the spatial query match statement.The validator in
src/Database/Validator/Queries.phpallows these four new spatial query types, but the match statement insrc/Database/Adapter/MariaDB.php(lines 1211-1228) does not handle them. These queries will throwUnknown spatial query methodat runtime when executed. The same issue exists insrc/Database/Adapter/Postgres.php.🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Adapter/MariaDB.php` around lines 1211 - 1228, The match in MariaDB.php handling spatial query methods misses Query::TYPE_COVERS, Query::TYPE_NOT_COVERS, Query::TYPE_SPATIAL_EQUALS and Query::TYPE_NOT_SPATIAL_EQUALS; add match arms that map TYPE_COVERS to "ST_Covers({$alias}.{$attribute}, {$geom})", TYPE_NOT_COVERS to "NOT ST_Covers(...)", TYPE_SPATIAL_EQUALS to "ST_SpatialEquals({$alias}.{$attribute}, {$geom})" and TYPE_NOT_SPATIAL_EQUALS to "NOT ST_SpatialEquals(...)" (use the same pattern/templating as the existing arms such as Query::TYPE_CONTAINS and Query::TYPE_EQUAL), and apply the identical changes to the corresponding match in Postgres.php so both adapters support the four new Query types.src/Database/Adapter/Mongo.php-1064-1069 (1)
1064-1069:⚠️ Potential issue | 🟠 MajorResolve partial-index types by attribute name, not loop offset.
Line 1165 builds
$indexAttributeTypesas[$attrName => $type], but Line 1067 reads$indexAttributeTypes[$i]. That falls back tostringfor numeric/date fields and can exclude the real field type from the partial index entirely.Suggested fix
if (in_array($type, [IndexType::Unique, IndexType::Key])) { $partialFilter = []; foreach ($attributes as $i => $attr) { - $attrType = $indexAttributeTypes[$i] ?? ColumnType::String->value; // Default to string if type not provided + $sourceAttribute = $index->attributes[$i] ?? $attr; + $attrType = $indexAttributeTypes[$sourceAttribute] ?? ColumnType::String->value; $attrType = $this->getMongoTypeCode($attrType); $partialFilter[$attr] = ['$exists' => true, '$type' => $attrType]; }🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Adapter/Mongo.php` around lines 1064 - 1069, The partial index builder in Mongo.php is using the loop index to look up attribute types but $indexAttributeTypes is keyed by attribute name; update the lookup in the loop that constructs $partialFilter (in the block handling IndexType::Unique and IndexType::Key) to use the attribute name ($attr) as the key instead of $i, e.g. fetch the type with $indexAttributeTypes[$attr] ?? ColumnType::String->value and then pass that into $this->getMongoTypeCode so the correct field type is used for the partial index filter.
🟡 Minor comments (12)
docker-compose.yml-6-7 (1)
6-7:⚠️ Potential issue | 🟡 MinorDocument local development setup requirement for sibling
queryrepository.The volume mount
../query/src:/usr/src/code/vendor/utopia-php/query/srcand build context..assume a specific local directory structure where thequeryrepository exists as a sibling. This dependency is not documented in CONTRIBUTING.md. While the volume mount failure is non-fatal in CI (the container starts regardless), local development withdocker compose up -d --buildwill fail without the siblingqueryrepository, making setup unclear for new contributors.Also applies to: 21-21
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@docker-compose.yml` around lines 6 - 7, The docker-compose configuration uses a sibling repo via the volume mount '../query/src:/usr/src/code/vendor/utopia-php/query/src' and a build context '..' (with dockerfile database/Dockerfile), which must be documented; update CONTRIBUTING.md to state that the `query` repository must exist as a sibling (or provide steps to clone it) before running `docker compose up -d --build`, note the mount is optional for CI and point to the alternative (remove/disable the volume or use a docker-compose.override.yml) for contributors who cannot have the sibling repo, and mirror this documentation for the other occurrence referenced (line 21) so local dev setup is explicit.README.md-636-638 (1)
636-638:⚠️ Potential issue | 🟡 MinorAdd import statement for
ForeignKeyActionin the example.The example shows
ForeignKeyAction::Cascade->valuebut doesn't include the necessaryusestatement. This could confuse developers trying to use the code.📝 Suggested addition before line 636
use Utopia\Database\ForeignKeyAction;🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@README.md` around lines 636 - 638, Add a use/import for the ForeignKeyAction enum in the example so the symbols like ForeignKeyAction::Cascade->value, ForeignKeyAction::SetNull->value, and ForeignKeyAction::Restrict->value resolve; specifically, add the statement to import the class (e.g., use Utopia\Database\ForeignKeyAction;) near the top of the snippet where the example code begins so ForeignKeyAction is available in that example context.README.md-758-776 (1)
758-776:⚠️ Potential issue | 🟡 MinorAdd import statement for
SetTypein the example.Similar to
ForeignKeyAction, theSetTypeenum usage examples should include the import statement for clarity.📝 Suggested addition before line 758
use Utopia\Database\SetType;🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@README.md` around lines 758 - 776, Add the missing import for the SetType enum used in the README examples by inserting the statement importing SetType (e.g., use Utopia\Database\SetType;) near the other example imports so the usages of SetType::Assign / SetType::Append / SetType::Prepend in the setAttribute examples resolve; ensure it appears alongside existing imports like Permission and Role so the sample snippet is self-contained and clear.src/Database/Relationship.php-21-31 (1)
21-31:⚠️ Potential issue | 🟡 Minor
toDocument()omitscollectionandkeyproperties.The
toDocument()method doesn't serialize$this->collectionor$this->key, butfromDocument()accepts both as inputs (collection as parameter, key from attribute). This asymmetry may cause data loss during round-trip serialization if these fields need to be preserved.If this is intentional (because these fields come from external context), consider documenting this behavior.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Relationship.php` around lines 21 - 31, The toDocument() method on Relationship currently omits $this->collection and $this->key while fromDocument() (and the static fromDocument method) accept collection and key, causing asymmetrical serialization; modify Relationship::toDocument() to include 'collection' => $this->collection and 'key' => $this->key in the returned Document (or, if omission is intentional, add a clear docblock on Relationship::toDocument() and fromDocument() explaining that collection and key are provided externally and therefore not persisted) so round-trip serialization is consistent.src/Database/Hook/RelationshipHandler.php-243-243 (1)
243-243:⚠️ Potential issue | 🟡 MinorUnused loop variable
$index.The
$indexvariable from the foreach loop is declared but never used. This was correctly flagged by static analysis.🔧 Proposed fix
- foreach ($relationships as $index => $relationship) { + foreach ($relationships as $relationship) {🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Hook/RelationshipHandler.php` at line 243, The foreach in RelationshipHandler.php declares an unused loop variable $index; change the loop signature from foreach ($relationships as $index => $relationship) to foreach ($relationships as $relationship) (or use foreach ($relationships as $_ => $relationship) if you prefer an explicit unused key) so the unused $index is removed, keeping the loop body using $relationship unchanged.src/Database/Hook/RelationshipHandler.php-1803-1814 (1)
1803-1814:⚠️ Potential issue | 🟡 MinorSame shadowing issue:
$documentin deleteCascade.The loop variable
$documentshadows the method parameter, same issue as indeleteSetNull. The loop iterates over junction documents while the original parameter is the document being deleted.🔧 Rename loop variable
- foreach ($junctions as $document) { + foreach ($junctions as $junctionDoc) { if ($side === RelationSide::Parent->value) { $this->db->deleteDocument( $relatedCollection->getId(), - $document->getAttribute($key) + $junctionDoc->getAttribute($key) ); } $this->db->deleteDocument( $junction, - $document->getId() + $junctionDoc->getId() ); }🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Hook/RelationshipHandler.php` around lines 1803 - 1814, In deleteCascade, the foreach loop uses $document which shadows the method parameter $document; rename the loop variable (e.g., to $junctionDocument or $junctionDoc) wherever it's used within the loop so the parameter $document remains the original document being deleted, and update the two deleteDocument calls to reference the renamed loop variable ($junctionDocument->getAttribute($key) and $junctionDocument->getId()) while leaving the method parameter untouched.src/Database/Attribute.php-76-92 (1)
76-92:⚠️ Potential issue | 🟡 Minor
fromArray()omitsstatusandoptionsfields.Unlike
fromDocument(), thefromArray()method doesn't extractstatusandoptionsfrom the input array. This asymmetry could cause data loss when attributes with these fields are serialized viatoDocument()and then reconstructed viafromArray().🔧 Proposed fix to include status and options
public static function fromArray(array $data): self { $type = $data['type'] ?? 'string'; return new self( key: $data['$id'] ?? $data['key'] ?? '', type: $type instanceof ColumnType ? $type : ColumnType::from($type), size: $data['size'] ?? 0, required: $data['required'] ?? false, default: $data['default'] ?? null, signed: $data['signed'] ?? true, array: $data['array'] ?? false, format: $data['format'] ?? null, formatOptions: $data['formatOptions'] ?? [], filters: $data['filters'] ?? [], + status: $data['status'] ?? null, + options: $data['options'] ?? null, ); }🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Attribute.php` around lines 76 - 92, fromArray() currently ignores the 'status' and 'options' keys so reconstructing an Attribute via Attribute::fromArray loses data present in toDocument()/fromDocument(); update Attribute::fromArray to read 'status' and 'options' from the input array (e.g. $data['status'] ?? <sensible default> and $data['options'] ?? <default empty array>) and pass them into the Attribute constructor (matching how fromDocument() does), ensuring defaults mirror fromDocument/toDocument behavior.src/Database/Hook/RelationshipHandler.php-272-283 (1)
272-283:⚠️ Potential issue | 🟡 MinorLoose equality comparison may cause unexpected behavior.
Using
==for comparing$oldValueand$value(Line 272) can lead to unexpected type coercion, especially when comparing Documents, arrays, or mixed types. This could cause relationship updates to be incorrectly skipped.🔧 Consider strict comparison or dedicated comparison logic
- if ($oldValue == $value) { + if ($this->areRelationshipValuesEqual($oldValue, $value)) {Alternatively, if loose comparison is intentional for this specific use case, add a comment explaining why.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Hook/RelationshipHandler.php` around lines 272 - 283, The loose comparison using $oldValue == $value in RelationshipHandler (around the block handling RelationType and RelationSide) can produce incorrect skips; change this to strict comparison or explicit comparison logic: use === for scalar checks, if both values are Document instances compare their IDs (e.g., $oldValue->getId() === $value->getId()), and for arrays use a deep equality check (or serialize/sort then compare) before deciding to set or remove attributes via $document->setAttribute/$document->removeAttribute; if loose comparison was intentional, replace the == with a clear comment explaining why and what cases rely on coercion.src/Database/Hook/RelationshipHandler.php-1727-1732 (1)
1727-1732:⚠️ Potential issue | 🟡 MinorParameter shadowing:
$documentreused in loop.The foreach loop variable
$documentshadows the method parameter$document(from Line 1640), which could cause confusion or unintended behavior if the original document is needed after the loop.🔧 Rename loop variable
- foreach ($junctions as $document) { - $this->db->skipRelationships(fn () => $this->db->deleteDocument( - $junction, - $document->getId() - )); + foreach ($junctions as $junctionDoc) { + $this->db->skipRelationships(fn () => $this->db->deleteDocument( + $junction, + $junctionDoc->getId() + )); }🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Hook/RelationshipHandler.php` around lines 1727 - 1732, The foreach loop reuses the variable name $document which shadows the method parameter $document; rename the loop variable (e.g., $junctionDocument or $junctionItem) and update its usage inside the loop so the method parameter remains available unchanged, keeping the existing calls to $this->db->skipRelationships(fn () => $this->db->deleteDocument($junction, <loop-var>->getId())); ensure only the loop variable name is changed (no other logic altered).src/Database/Traits/Indexes.php-359-366 (1)
359-366:⚠️ Potential issue | 🟡 MinorFix rollback index TTL default: should be
0, not1, to match TTL semantics elsewhere.The rollback index uses
getAttribute('ttl', 1)as default, but throughout the codebase TTL semantics treat0(or missing attribute) as "no TTL configured". InDocuments.php:247, the code explicitly checksif ($ttlSeconds <= 0 || !$ttlAttr)to determine whether to skip TTL processing. Additionally, theValidator/Index.php:810usesgetAttribute('ttl', 0)as default. Using1as the default means if the original index had no TTL attribute, the rollback would incorrectly create an index with a 1-second TTL, causing documents to expire prematurely. Change the default to0to preserve the "no TTL" state when the attribute is missing.🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Traits/Indexes.php` around lines 359 - 366, The rollback index creation uses getAttribute('ttl', 1) which incorrectly sets a 1s TTL when the attribute is missing; change the default to 0 so $rollbackIndex is created with no TTL when absent by updating the call on $indexDeleted->getAttribute('ttl', 1) to use 0 (the code building the Index instance with IndexType::from(...) and attributes/lengths/orders should remain the same) to match Validator/Index.php and Documents.php TTL semantics.src/Database/Adapter/SQLite.php-43-70 (1)
43-70:⚠️ Potential issue | 🟡 MinorSQLite advertises it does not support
Capability::Upsertsbut still implements upsert functionality.Line 62 removes
Capability::Upsertsfrom the advertised capabilities, yet SQLite inheritsupsertDocuments()from SQL and overridesexecuteUpsertBatch()to handle SQLite's ON CONFLICT syntax. This creates an inconsistency: the capability advertisement does not match the runtime behavior. While this doesn't currently cause dead code (no code gate upsert calls onsupports(Capability::Upserts)), the mismatch violates the contract that removed capabilities should not be implemented. Either re-addCapability::Upsertsto SQLite's capabilities, or remove theexecuteUpsertBatch()override if SQLite should not support upserts.🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Adapter/SQLite.php` around lines 43 - 70, The SQLite adapter currently removes Capability::Upserts in capabilities() while still implementing upsertDocuments() and overriding executeUpsertBatch(), causing a contract mismatch; fix by either (A) re-adding support: remove Capability::Upserts from the $remove array in capabilities() so parent::capabilities() reports Upserts, or (B) disable runtime support: delete or revert the SQLite::executeUpsertBatch() override (and any SQLite-specific upsert helpers) so the adapter no longer implements upsert behavior — choose one approach and make the changes consistently (references: capabilities(), Capability::Upserts, upsertDocuments(), executeUpsertBatch(), parent::capabilities()).src/Database/Adapter/Mongo.php-1081-1081 (1)
1081-1081:⚠️ Potential issue | 🟡 MinorRemove the
->valueaccessor to fix the unreachable readiness loop.At line 1081,
$typeis anIndexTypeenum case (assigned from$index->typeat line 44), but the comparison$type === IndexType::Unique->valuecompares an enum case object to its backing scalar value. This will never match in PHP—enum cases and their backing values are distinct types. As a result, the readiness loop is unreachable and the code never waits for unique indexes to be fully built.The fix is to compare the enum case directly:
Fix
- if ($type === IndexType::Unique->value) { + if ($type === IndexType::Unique) {This is consistent with all other enum comparisons in the same function (lines 1020, 1042, 1050, 1055, 1060), which use enum cases without the
->valueaccessor.🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Adapter/Mongo.php` at line 1081, The readiness loop is unreachable because it compares the enum case object $type to the backing scalar of IndexType::Unique (using ->value); locate the comparison of $type and IndexType::Unique in the readiness/wait loop and remove the ->value accessor so the code compares the enum case directly (i.e., use IndexType::Unique) to match how other enum checks in this function handle $type.
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
Run ID: 3dfcd406-2dfe-4455-a2f0-baf7fa17d195
⛔ Files ignored due to path filters (1)
composer.lockis excluded by!**/*.lock
📒 Files selected for processing (128)
.github/workflows/tests.ymlDockerfileREADME.mdbin/tasks/relationships.phpcomposer.jsondocker-compose.ymlsrc/Database/Adapter.phpsrc/Database/Adapter/Feature/Attributes.phpsrc/Database/Adapter/Feature/Collections.phpsrc/Database/Adapter/Feature/ConnectionId.phpsrc/Database/Adapter/Feature/Databases.phpsrc/Database/Adapter/Feature/Documents.phpsrc/Database/Adapter/Feature/Indexes.phpsrc/Database/Adapter/Feature/InternalCasting.phpsrc/Database/Adapter/Feature/Relationships.phpsrc/Database/Adapter/Feature/SchemaAttributes.phpsrc/Database/Adapter/Feature/Spatial.phpsrc/Database/Adapter/Feature/Timeouts.phpsrc/Database/Adapter/Feature/Transactions.phpsrc/Database/Adapter/Feature/UTCCasting.phpsrc/Database/Adapter/Feature/Upserts.phpsrc/Database/Adapter/MariaDB.phpsrc/Database/Adapter/Mongo.phpsrc/Database/Adapter/Mongo/RetryClient.phpsrc/Database/Adapter/MySQL.phpsrc/Database/Adapter/Pool.phpsrc/Database/Adapter/Postgres.phpsrc/Database/Adapter/SQL.phpsrc/Database/Adapter/SQLite.phpsrc/Database/Attribute.phpsrc/Database/Capability.phpsrc/Database/CursorDirection.phpsrc/Database/Database.phpsrc/Database/Document.phpsrc/Database/Helpers/Permission.phpsrc/Database/Hook/MongoPermissionFilter.phpsrc/Database/Hook/MongoTenantFilter.phpsrc/Database/Hook/PermissionFilter.phpsrc/Database/Hook/PermissionWrite.phpsrc/Database/Hook/Read.phpsrc/Database/Hook/Relationship.phpsrc/Database/Hook/RelationshipHandler.phpsrc/Database/Hook/TenantFilter.phpsrc/Database/Hook/TenantWrite.phpsrc/Database/Hook/Write.phpsrc/Database/Hook/WriteContext.phpsrc/Database/Index.phpsrc/Database/Mirror.phpsrc/Database/Operator.phpsrc/Database/OperatorType.phpsrc/Database/OrderDirection.phpsrc/Database/PermissionType.phpsrc/Database/Query.phpsrc/Database/RelationSide.phpsrc/Database/RelationType.phpsrc/Database/Relationship.phpsrc/Database/SetType.phpsrc/Database/Traits/Attributes.phpsrc/Database/Traits/Collections.phpsrc/Database/Traits/Databases.phpsrc/Database/Traits/Documents.phpsrc/Database/Traits/Indexes.phpsrc/Database/Traits/Relationships.phpsrc/Database/Traits/Transactions.phpsrc/Database/Validator/Attribute.phpsrc/Database/Validator/Datetime.phpsrc/Database/Validator/Index.phpsrc/Database/Validator/IndexedQueries.phpsrc/Database/Validator/Operator.phpsrc/Database/Validator/Permissions.phpsrc/Database/Validator/Queries.phpsrc/Database/Validator/Queries/Document.phpsrc/Database/Validator/Queries/Documents.phpsrc/Database/Validator/Query/Filter.phpsrc/Database/Validator/Query/Limit.phpsrc/Database/Validator/Query/Offset.phpsrc/Database/Validator/Sequence.phpsrc/Database/Validator/Spatial.phpsrc/Database/Validator/Structure.phptests/e2e/Adapter/Base.phptests/e2e/Adapter/MariaDBTest.phptests/e2e/Adapter/MirrorTest.phptests/e2e/Adapter/MongoDBTest.phptests/e2e/Adapter/MySQLTest.phptests/e2e/Adapter/PoolTest.phptests/e2e/Adapter/PostgresTest.phptests/e2e/Adapter/SQLiteTest.phptests/e2e/Adapter/Schemaless/MongoDBTest.phptests/e2e/Adapter/Scopes/AttributeTests.phptests/e2e/Adapter/Scopes/CollectionTests.phptests/e2e/Adapter/Scopes/CustomDocumentTypeTests.phptests/e2e/Adapter/Scopes/DocumentTests.phptests/e2e/Adapter/Scopes/GeneralTests.phptests/e2e/Adapter/Scopes/IndexTests.phptests/e2e/Adapter/Scopes/ObjectAttributeTests.phptests/e2e/Adapter/Scopes/OperatorTests.phptests/e2e/Adapter/Scopes/PermissionTests.phptests/e2e/Adapter/Scopes/RelationshipTests.phptests/e2e/Adapter/Scopes/Relationships/ManyToManyTests.phptests/e2e/Adapter/Scopes/Relationships/ManyToOneTests.phptests/e2e/Adapter/Scopes/Relationships/OneToManyTests.phptests/e2e/Adapter/Scopes/Relationships/OneToOneTests.phptests/e2e/Adapter/Scopes/SchemalessTests.phptests/e2e/Adapter/Scopes/SpatialTests.phptests/e2e/Adapter/Scopes/VectorTests.phptests/e2e/Adapter/SharedTables/MariaDBTest.phptests/e2e/Adapter/SharedTables/MongoDBTest.phptests/e2e/Adapter/SharedTables/MySQLTest.phptests/e2e/Adapter/SharedTables/PostgresTest.phptests/e2e/Adapter/SharedTables/SQLiteTest.phptests/unit/DocumentTest.phptests/unit/OperatorTest.phptests/unit/PermissionTest.phptests/unit/QueryTest.phptests/unit/Validator/AttributeTest.phptests/unit/Validator/AuthorizationTest.phptests/unit/Validator/DocumentQueriesTest.phptests/unit/Validator/DocumentsQueriesTest.phptests/unit/Validator/IndexTest.phptests/unit/Validator/IndexedQueriesTest.phptests/unit/Validator/OperatorTest.phptests/unit/Validator/QueriesTest.phptests/unit/Validator/Query/FilterTest.phptests/unit/Validator/Query/OrderTest.phptests/unit/Validator/Query/SelectTest.phptests/unit/Validator/QueryTest.phptests/unit/Validator/SpatialTest.phptests/unit/Validator/StructureTest.php
There was a problem hiding this comment.
Actionable comments posted: 3
🧹 Nitpick comments (2)
src/Database/Adapter/Pool.php (1)
80-88: Use the short class name for consistency.Line 80 uses the fully qualified class name
\Utopia\Database\Capabilityeven thoughCapabilityis already imported at line 7. Use the short name for consistency with the rest of the file.Suggested fix
- public function supports(\Utopia\Database\Capability $feature): bool + public function supports(Capability $feature): bool { return $this->delegate(__FUNCTION__, \func_get_args()); }🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Adapter/Pool.php` around lines 80 - 88, Change the fully-qualified parameter type in the supports method to the short imported class name: replace "\Utopia\Database\Capability $feature" with "Capability $feature" in the supports method signature so it matches the rest of the file and uses the imported Capability symbol; ensure the method still calls $this->delegate(__FUNCTION__, func_get_args()) unchanged.src/Database/Adapter/SQLite.php (1)
247-248: Remove the duplicated_permsindex creation.These same two indexes are already created on Lines 247-248. Repeating them here just adds extra catalog probes and will become observable if index-create hooks/events are added later.
♻️ Proposed cleanup
- $this->createIndex("{$id}_perms", new Index(key: '_index_1', type: IndexType::Unique, attributes: ['_document', '_type', '_permission'])); - $this->createIndex("{$id}_perms", new Index(key: '_index_2', type: IndexType::Key, attributes: ['_permission', '_type']));Also applies to: 265-266
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/Database/Adapter/SQLite.php` around lines 247 - 248, The code calls $this->createIndex twice for the same "{$id}_perms" table with identical Index(...) definitions (using Index(key: '_index_1', type: IndexType::Unique, attributes: ['_document','_type','_permission']) and Index(key: '_index_2', type: IndexType::Key, attributes: ['_permission','_type'])), causing duplicate index creation; remove the duplicated createIndex calls (the repeated "{$id}_perms" invocations) so each index is created only once, leaving a single call to createIndex for each Index definition (references: createIndex, Index, IndexType).
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@src/Database/Adapter/SQLite.php`:
- Around line 45-65: The capabilities() method currently removes
Capability::Upserts from the $remove list which causes
supports(Capability::Upserts) to return false even though this class implements
SQLite-specific ON CONFLICT expressions and executeUpsertBatch(); fix by
removing Capability::Upserts from the $remove array (so Upserts is reported as
supported) and ensure any related logic in executeUpsertBatch() and
conflict-expression generation remains intact to be exercised when
supports(Capability::Upserts) is checked.
In `@src/Database/Mirror.php`:
- Around line 1074-1086: The setRelationshipHook method in Mirror currently
discards the caller's RelationshipHook by always installing a new
RelationshipHandler on the source/destination; instead, forward the provided
$hook (or null) directly to the underlying databases. In the
Mirror::setRelationshipHook(?RelationshipHook $hook) implementation replace the
calls that pass new RelationshipHandler($this->source)/new
RelationshipHandler($this->destination) with the original $hook (or null) so the
caller's custom RelationshipHook is preserved when calling
$this->source->setRelationshipHook(...) and
$this->destination->setRelationshipHook(...).
- Around line 321-333: The code calls Attribute::fromDocument($document) after
running filters in $this->writeFilters, but Filter::beforeCreateAttribute() can
return null to indicate the attribute should be skipped; update the blocks that
call beforeCreateAttribute (e.g., the loop using $this->writeFilters and the
code that assigns $filteredAttribute) to check if $document is null after the
loop and, if so, skip further processing (do not call Attribute::fromDocument) —
apply the same null-check/skipping logic to the other symmetric block that
processes attributes (the second occurrence mentioned around the attribute
handling).
---
Nitpick comments:
In `@src/Database/Adapter/Pool.php`:
- Around line 80-88: Change the fully-qualified parameter type in the supports
method to the short imported class name: replace "\Utopia\Database\Capability
$feature" with "Capability $feature" in the supports method signature so it
matches the rest of the file and uses the imported Capability symbol; ensure the
method still calls $this->delegate(__FUNCTION__, func_get_args()) unchanged.
In `@src/Database/Adapter/SQLite.php`:
- Around line 247-248: The code calls $this->createIndex twice for the same
"{$id}_perms" table with identical Index(...) definitions (using Index(key:
'_index_1', type: IndexType::Unique, attributes:
['_document','_type','_permission']) and Index(key: '_index_2', type:
IndexType::Key, attributes: ['_permission','_type'])), causing duplicate index
creation; remove the duplicated createIndex calls (the repeated "{$id}_perms"
invocations) so each index is created only once, leaving a single call to
createIndex for each Index definition (references: createIndex, Index,
IndexType).
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
Run ID: 6220a957-b838-4386-98ab-e1a755c06239
📒 Files selected for processing (7)
.gitignoresrc/Database/Adapter/Pool.phpsrc/Database/Adapter/SQLite.phpsrc/Database/Database.phpsrc/Database/Mirror.phptests/e2e/Adapter/Scopes/Relationships/ManyToManyTests.phptests/e2e/Adapter/Scopes/SchemalessTests.php
✅ Files skipped from review due to trivial changes (1)
- .gitignore
6404da4 to
4f96135
Compare
|
@coderabbitai full review |
✅ Actions performedFull review triggered. |
263bfd4 to
a07ec0e
Compare
d38a772 to
fa461e5
Compare
There was a problem hiding this comment.
Code Review
Reviewed this PR for CRITICAL/HIGH/MEDIUM findings (bugs and CLAUDE.md compliance).
1 HIGH finding — a permission-check regex regression that was silently reintroduced during the query-lib refactor. Details inline.
Other items considered and cleared:
- Import ordering in
Hook/Relationships.phpandORM/EntityMapper.php— conforms to Pint/PSR-12 case-insensitive ordering. - Missing
readonlyon#[Attribute]mapping classes — AGENTS.md phrases this as a conditional preference ("readonly properties for immutable data"), not a hard rule; not a clear violation. - Bug scan across
SQL.php, adapter files, traits, hooks, validators, andORM/— no high-confidence bugs found in the new code (several initially-suspicious patterns turned out to be pre-existing onmainor safely handled).
|
Claude pushed fixes from: improvement |
There was a problem hiding this comment.
Code Review
Reviewed this PR for CRITICAL/HIGH/MEDIUM findings (bugs and CLAUDE.md compliance). 1 HIGH finding — a cross-tenant data-leak regression in the newly extracted TenantFilter hook. Details inline.
Other items considered and cleared:
- Duplicate
createIndexcalls inSQLite.php:267-268— no-op becausecreateIndexhas an explicitIF NOT EXISTSshort-circuit (lines 438-449). Redundant but not a bug; CodeRabbit already flagged as a nit. Mirror::addHookconstructing freshRelationships($this->source/destination)— the new hook is bound to the correct target database; not a clear regression.- Missing
readonlyon newAttribute/Index/Collection/Relationshipvalue objects — AGENTS.md phrases this as a conditional preference ("readonly properties for immutable data"), not an enforced rule; the prior review cycle on this PR already dismissed the analogous ORM case. - Snake_case
test_*method naming across ~30 new test files — widespread deliberate convention change for the refactor, not a discrete bug. - Bug scan across new
Hook/**, adapters, traits, and validators — no high-confidence regressions found beyond the one flagged below.
|
Claude pushed fixes from: improvement |
CI healing — run 24884705670Root-caused two undefined-constant errors introduced by the recent merge from 1. Unit Tests (3 errors) —
|
|
Claude pushed fixes from: healing |
CI heal (run 24886550663)Failure: Root cause: After merging Fix: Route through the query builder's - $result = $builder->insert();
+ $result = $this->skipDuplicates ? $builder->insertOrIgnore() : $builder->insert(); |
There was a problem hiding this comment.
Code Review — critical & warning findings fixed
Ran /skills:improve on the query-library core of this PR and resolved 7 CRITICAL / 4 WARNING findings across the recent-delta files. Full diff is large (340 files), so the review was scoped to Query.php, Validator/Query/*, and Validator/Queries.php — the PR's core extraction theme.
Dimension summary
- Security / correctness:
ColumnType::from()\ValueError,Filterswitch missing 10ColumnTypecases,Order::addAggregationAliasesschema mutation leak across requests, aggregate-method dispatch missing half ofMethod::isAggregate()(stddevPop, varPop, bitAnd/Or/Xor, …),toArray()crash on non-Query child ofAnd/Or/ElemMatch, pre-existing fatal-on-load from incompatibleLOGICAL_TYPESvisibility andparse()signature drift with the base library. - Performance / maintainability: dead duplicate-schema-lookup block in
Filter::isValidAttributeAndValues, nested-recursion overUnion/UnionAllthat are not filter containers, drift risk betweenshape()andtoArray()using two different literals for the same set.
Test result delta
Before this commit, the branch could not load Utopia\Database\Query (fatal: narrowed visibility on LOGICAL_TYPES; incompatible parse() signature). Fixing those plus the groupForDatabase order-attributes regression dropped unit-test failures from 31 errors + 1 failure → 2 errors + 1 failure. The 3 remaining failures are environmental (PHPUnit Swoole class, xdebug + PHP 8.2 dynamic-property on DateMalformedStringException) and are not caused by this PR.
Not addressed (intentional — out of this cycle's scope)
- DRY refactor of the 5 near-identical
Aggregate/Distinct/GroupBy/Having/Joinvalidator skeletons. - Single-pass restructure of
Queries::isValid(currently parses strings twice — correctness-neutral). - Hoisting the per-value
switch-built validator inFilter::isValidAttributeAndValuesout of the foreach (performance-only). - Cycle-guard / node-count cap on
Query::shape()(theoretical DoS — unlikely to be reachable in practice).
These are worth doing but are architectural and carry regression risk for a follow-up PR.
|
Claude pushed fixes from: improvement |
CI heal — follow-up (after local verification)My first fix touched
Local verification: Files touched:
|
|
Claude pushed fixes from: healing |
CI failure — root cause and fixFailed run: #24887417627 What failed Every adapter test job errored with:
originating from Why Those three cases were added upstream in Fix Ran |
|
Claude pushed fixes from: healing |
Every getDocument() inside a transaction skipped the document cache, collection definitions included, so each write re-read its own collection definition from _metadata inside its transaction: +1 statement per update, increase, decrease and delete, +3 in withTransaction(update + get), and +5 to +45 per nested relationship create (B-03, J3C-003/J3C-014). 7.3.12 served hits there. A read inside a transaction now serves a cached copy when the invalidation scope of its coroutine started the transaction and has not written that document. Single-document writes record their lower-cased document key in the scope (transactionWrites), so a document the transaction wrote, under any casing an adapter may match, is read from the adapter: another reader can refill its slot with the committed row after the purge inside the transaction. Batch writes and schema changes keep the collection's epoch blocked, which already makes those reads miss. A transaction the database did not start (one begun on the adapter, or a Pool transaction another coroutine owns) keeps reading uncached, since its earlier writes may sit in scopes that have already closed. Reads inside a transaction still never fill: a transaction reads its snapshot, which under REPEATABLE READ can predate another writer's commit and post-commit purge, and a lease taken after that purge would accept the stale row. Collection definitions no longer split their cache field by whether the relationship hook is enabled: a definition has no relationship attributes, so the hook never changes what is read. Without this, every lookup made under skipRelationships() inside a nested create missed a field that only a read outside a transaction fills. The cost moves from the database to the cache: the locking read's collection lookup is now a cache hit (6 round trips until one round trip per lookup), so the single-document writes that read first take 14 round trips on a warm cache (8 before), pinned in DocumentCacheInvalidationTest. Co-Authored-By: Claude Opus 5.5 <[email protected]>
A getDocument() miss re-read the row with authorization skipped before
publishing the shared negative entry, in case the adapter had filtered it out
by the caller's permissions. Collection definitions paid that re-read too, so
getCollection('missing') and createCollection()'s free-id check cost two
_metadata reads where 7.3.12 paid one (H-07/Q10, JDB13-015; J3C-025, the two
extra _metadata reads of schema.create_collection_with_10_attributes).
No adapter filters definitions by permission: every write resolves its
collection through them, so a filtered definition would already break writes
for callers without read access. The re-read stays for every other collection,
where MongoFilteredMissCacheTest still relies on it.
The other half of J3C-025, the cached miss retired by an unrelated _metadata
write, went with the per-document purges of j-db-fix-04;
testCreateCollectionAfterAProbeReadsNoDefinition pins it.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
ReadWritePool sends reads outside the sticky window to a replica, and getDocument() and find() cached whatever came back. A lagging replica's row, its "not found", or its query result then went into the shared document and query caches under the current epoch, so every handle kept reading the old version until the next write or the 24 h TTL (H-02, J4L-006). ReadWritePool now records, per coroutine, whether the latest data read was served by the replica pool, and Database skips the document save, the negative marker and the query-cache save for such a read. The record lives in the coroutine's context (an instance field outside coroutines), so a read on one coroutine cannot hide or fake where another coroutine's read was served, and it is freed with the coroutine. Reads the pool sends to the primary (inside a transaction, in the sticky window, or H-01's write-deciding reads) fill as before, so the caches are still filled right after writes; a replica-served read is simply read again from the replica next time. Co-Authored-By: Claude Opus 5.5 <[email protected]>
| $collect = function (Document $updated) use (&$cleared): void { | ||
| $cleared[] = $updated; | ||
| }; |
There was a problem hiding this comment.
Unused Reports Retain Documents
For a SetNull delete with many related documents, this callback keeps every updated document across all chunks, even when there are no lifecycle listeners and the report will be discarded. This removes the memory bound provided by chunking and can increase a worker’s memory use substantially. Collect the documents only when a report is needed.
Prompt To Fix With AI
This is a comment left during a code review.
Path: src/Database/Hook/Relationships.php
Line: 2071-2073
Comment:
**Unused Reports Retain Documents**
For a SetNull delete with many related documents, this callback keeps every updated document across all chunks, even when there are no lifecycle listeners and the report will be discarded. This removes the memory bound provided by chunking and can increase a worker’s memory use substantially. Collect the documents only when a report is needed.
---
For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.…e per coroutine Relationship population reads chunks of related ids in Promise::map() child coroutines. The children shared the handle's Authorization status and the relationship hook's flags, and find(), skip() and skipRelationships() saved, changed and restored that shared state, so siblings restored each other's values in completion order. One find() could leave the handle with authorization and relationships disabled, after which reads returned documents that only document-level grants protect. silent() was keyed per coroutine id, so the children also escaped the caller's silence (B-11). Scoped toggles now open an override that belongs to the calling coroutine and the coroutines it starts (State\Value): Authorization::skip()/withStatus(), find()'s authorization skip (now the scoped form), skipRelationships(), skipRelationshipsExistCheck(), the batch-population flag and silent(). A read takes the innermost override of the current coroutine, else the nearest ancestor's, else the handle-wide value. Unscoped writes (setStatus, enable, disable, reset, setEnabled, setCheckExist) change the current coroutine's innermost override when there is one and the handle-wide value otherwise, so a disable() made in a worker or migration coroutine is still seen by the coroutines it starts and by later work, as appwrite relies on. Chunks are read concurrently only inside a coroutine, on Adapter\Pool (one borrowed connection per read) and outside a transaction; otherwise they are read one after another in the caller. The Pool keeps its transaction pin per handle. Each concurrent read starts from Database::snapshot() of its caller through withSnapshot(), which Mirror can reuse to run replication under the caller's state. Chunk results are merged in chunk order. Gaps: J4C-006 (alias J4A-007), B-11. Co-Authored-By: Claude Opus 5.5 <[email protected]>
utopia-php/query 0.6.1 compiles contains/containsAny/containsAll/notContains on SQLite array attributes as `json_each.value = json(?)` with the JSON encoding bound. json() returns JSON text (`"a"`, `1`, `true`) while json_each.value holds the element's SQL value (a, INTEGER 1, 1), so no string, integer, double or boolean element ever matched: containsAny and containsAll returned no rows and notContains returned every row (J4C-009 = J4Q-001, critical). Compare with json_extract(?, '$'), which turns the bound JSON into the same SQL value json_each yields (verified for strings, non-ASCII, quotes, integers, doubles and booleans in the j-database-09 probes). The override sits on compileJsonContainsExpr()/compileJsonOverlapsExpr(), so the query-builder JSON filters (jsonContains/jsonNotContains/jsonOverlaps) are fixed as well, not only the array filter. notContains and jsonNotContains now also require the column to be non-NULL: MariaDB/MySQL (NOT JSON_OVERLAPS/JSON_CONTAINS), PostgreSQL (NOT @>), Memory and Redis all exclude a document whose array is NULL, while `NOT EXISTS (... json_each(NULL))` kept it on SQLite (user decision W1-11). Co-Authored-By: Claude Opus 5.5 <[email protected]>
notContains on an array attribute compiled to `{attr: {$nin: [...]}}`, and
MongoDB's $nin also matches documents where the field is missing or null.
MariaDB, MySQL, PostgreSQL, Memory, Redis and (now) SQLite exclude those
documents, so the same query returned extra rows on MongoDB only. Adding
`$ne: null` keeps empty arrays and drops missing/null ones, matching the
other engines (user decision W1-11: notContains excludes NULL arrays on
every engine).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
SQLite dropped Capability::QueryContains, so testFindContains, testFindNotContains and the contains block of testArrayAttribute skipped on both SQLite lanes and nothing caught array contains returning wrong rows there (J3-001). With element comparison fixed, all three gated tests pass on SQLite, plain and shared tables, so the capability is declared again. testArrayContainsQueriesOnScalarArrays pins one data set (strings with non-ASCII and quote characters, a numeric string, integers, doubles, booleans, an empty array and a NULL array) and the same expected rows for containsAny, containsAll, notContains and the deprecated contains form on every adapter with defined attributes, so all lanes have to agree, including on notContains leaving out the NULL array. Co-Authored-By: Claude Opus 5.5 <[email protected]>
find() results were stored under one key per epoch and query (`<scope>#<epoch>:<query>`). Every invalidation published a new epoch and the old keys stayed in Redis with no expiry, so the key count grew by one per invalidate-and-fill cycle (J4C-004, query half of B-05-R). Results now live in one hash per collection scope (the scope key itself), one field per query. The value records the epoch it was filled under and its own field; a hit needs both to match, so a fill that raced an invalidation is never served under the next epoch, and caches that ignore fields (Memory, Filesystem) never serve one query another query's rows (W1-23): there a scope holds one result at a time. Blocking a collection purges the hash, so a rotation clears the previous epoch's fields and, on leasable caches, also bumps the lease generation for the whole scope (W1-20: accepted; the lease and tombstone are collection-wide and a collection's results share one key/slot). Round trips are unchanged: a hit is 3, a miss 5. The Greptile r4120070321 thread asked for no exact key set: the rewritten test pins the key and field counts after cycle 20 to cycle 1's. Co-Authored-By: Claude Opus 5.5 <[email protected]>
…'s transaction deleteDocument() read the document inside the withMutation() closure and leaked it out through a by-reference capture, so after the transaction PHPStan only knew it as ?Document. The old triggerHooks() took mixed and hid that; the typed triggerDeleteHooks() added for the peer updates does not. The closure now returns the document it deleted, or null when there was nothing to delete or the adapter deleted nothing, so the hooks receive a Document by construction and the by-reference capture is gone. The e2e test read hook payloads as list<mixed> and passed them to closures typed on Document. EventRecorder::getDocuments() returns the payloads of an event as list<Document> and throws on anything else, so a wrong payload fails the test loudly instead of being cast, and the id closure takes Document variadics so its input is typed natively. Co-Authored-By: Claude Opus 5.5 <[email protected]>
…eries() reports cache failures purgeCachedCollection() only advanced the document-cache epoch, so a collection's find() results stayed cached after it: a write made around the query cache (a migration, another process without it) stayed invisible until the next write through the library (H-08). It now also invalidates the collection's query cache, as purgeCachedQueries() does; a failure propagates like the document half's, and the callers retry it. purgeCachedQueries() loaded the withCache() region epoch outside its try, so a cache failure threw although UPGRADE documents a `false` return. The load, purge and save of that epoch now log a warning and make it return false. Co-Authored-By: Claude Opus 5.5 <[email protected]>
With a query cache installed, a cache backend error anywhere in find()'s query-cache block (the epoch load, the generation reads, the entry load, the fill) made find() throw, while getDocument() logs a warning and reads the database (H-09). find() now does the same: a failure while resolving or loading an entry reads uncached and fills nothing; a failure while filling keeps the rows it read. Building the cache key stays outside the catch, so an invalid query still throws. QueryCache::getEpoch() threw on an epoch value it did not write; it is now a miss for that read (uncached, no fill), and the next invalidation overwrites it. Co-Authored-By: Claude Opus 5.5 <[email protected]>
| $this->assertSame(['value' => 'first'], $first); | ||
| $this->assertSame(['value' => 'second'], $second); | ||
| $this->assertSame(['value' => 'first'], $cachedFirst); | ||
| $this->assertSame(1, $firstCalls); | ||
| $this->assertSame(1, $secondCalls); | ||
| $this->assertSame(['first-field', 'second-field'], $cache->list('key')); | ||
| $this->assertSame(3, $cache->getSize()); |
There was a problem hiding this comment.
This assertion requires the cache to use exactly three storage keys for two hash fields. A change to the cache layout could preserve the correct results but fail the test. The assertions above already verify both cached values and that neither callback runs again. The repository requires tests to check observable behavior rather than internal storage details; this requirement must be satisfied before merging.
| $this->assertSame(3, $cache->getSize()); |
Context Used: Call out and harshly judge implementation-coupled tests. We don't mirror source code, configuration, or version pins in assertions. We test observable behavior; use linters for syntax and schema checks. (source)
Prompt To Fix With AI
This is a comment left during a code review.
Path: tests/unit/QueryCacheTest.php
Line: 205
Comment:
**Test Pins Cache Key Count**
This assertion requires the cache to use exactly three storage keys for two hash fields. A change to the cache layout could preserve the correct results but fail the test. The assertions above already verify both cached values and that neither callback runs again. The repository requires tests to check observable behavior rather than internal storage details; this requirement must be satisfied before merging.
```suggestion
```
**Context Used:** Call out and harshly judge implementation-coupled tests. We don't mirror source code, configuration, or version pins in assertions. We test observable behavior; use linters for syntax and schema checks. ([source](https://app.greptile.com/review/custom-context?memory=instruction-0))
---
For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!
A writer killed between its tombstone and its activation left #started ahead of #finished for good. The tombstone only lapsed with the region TTL once the scope was quiescent, so the collection's query cache stayed off until a flush, and every later write left its own tombstone behind because its activation saw the dead writer still counted in flight (PN-db03-2, the query-cache twin of B-01). Bound: a new writer timeout (QueryCache constructor, default 3600 s). - Read side: a tombstone older than the writer timeout lapses even while writes are counted in flight (older than the region TTL once none are, as before). The lapsed epoch is derived from the tombstone and the finished generation instead of the constant initial epoch, so a pre-block fill that raced the block is never re-exposed, and any later activation retires what readers filled during the lapse. - Reconciliation: tokens now record their creation time. An activation that finishes its own write while others are still counted lists the scope's owner registrations; when every other one is older than the writer timeout it releases them and publishes a fresh epoch stamped with the current started generation. A token without a time counts as live, and caches that keep no fields (no listing) skip this and rely on the read side. - A writer whose registration was released that way still publishes a fresh epoch when it activates, so what readers filled during its transaction is retired. A live transaction younger than the writer timeout keeps its block, as before. One older than it is treated as abandoned: readers may fill from the pre-commit state while it runs, and its activation retires those fills, so only the window between its commit and its activation can serve them. Co-Authored-By: Claude Opus 5.5 <[email protected]>
…ion was lost When the server ended the session inside an outer withTransaction(), the nested call's SAVEPOINT reached the dead connection, the PDO wrapper reconnected and rethrew, and rolling back on the new connection failed. withTransaction() then zeroed the counter and retried, so the nested call began a fresh top-level transaction and committed its work alone; the outer commit saw no transaction and returned false, and the outer call returned normally with its own writes lost. withTransaction() now records the depth it was entered at. A nested call whose rollback leaves the counter below that depth, or a call whose own level vanished underneath it, throws a TransactionException instead of retrying, so no call begins a fresh top-level transaction inside what its caller still treats as an open one, and the failure reaches the outer call. The check runs before the non-retryable exception list, so a duplicate the caller expects to catch cannot mask the lost transaction. Nested retries while the enclosing transaction holds, and the top-level reset-and-retry after a failed rollback (#898), keep their behaviour. SQL::commitTransaction() throws when a nested commit finds the driver no longer holds the transaction, instead of resetting the counter and returning false. The top-level case still returns false: #898's testWithTransactionResetsCounterOnRollbackFailureDuringRetry stubs inTransaction() as false for a transaction it expects to commit. Closes J4L-001 (ledger 14.P1). Co-Authored-By: Claude Opus 5.5 <[email protected]>
UPGRADE.md promises document_purge once the write's transaction has committed, but a write inside a caller's withTransaction() fired it inside that transaction, before the commit, once per retried attempt, and even when the caller's transaction later rolled back. A listener that purges its own cache could then refill it from the uncommitted state or purge for a write that never happened. Writes now queue the event in the invalidation scope from inside their mutation. The outermost scope fires the queue after its commit and the post-commit invalidation; a rollback drops it, every retried attempt starts from the events queued before its transaction, and a nested transaction that fails drops the events queued inside it. A queued event keeps the tenant and the hook silences in force when the document was written. A write the adapter holds no transaction for (MongoDB without a replica set) is already durable and still announces at once. Top-level writes fire once after their commit, as before. Closes JDB13-019 (H-14). Co-Authored-By: Claude Opus 5.5 <[email protected]>
…ion fails A cache failure in the invalidation after a commit made updateDocument(), deleteDocument() and the other writes throw with the data committed, and document_purge never fired: the triggers ran only after withMutation() returned. Listeners that keep their own caches were left serving the pre-write state until their TTL. The outermost scope now fires the queued document_purge events after the post-commit invalidation whether or not it succeeded, and then rethrows the invalidation failure, which still reaches the caller first (the throw stays, as DocumentCacheEpochTest::testActivationPropagatesAnOwnerReleaseFailure pins). Every queued event is delivered even if a listener throws; the first failure is rethrown after the last event. Closes JDB13-014 (H-06). Co-Authored-By: Claude Opus 5.5 <[email protected]>
…tion
Every cache lookup read #started, #finished, #epoch and both generations
again before the entry: 6 round trips per lookup, 12 for a getDocument()
hit whose collection lookup paid the same (bench-core B-04 read path,
J3C-001; the reads 1.2-4x slower than 7.3.12, J3C-002).
After j-db-fix-04 a definition is invalidated per document, so its cache
entry can carry the epoch its collection's documents are cached under:
getCollection() is one round trip and getDocument() two, as in 7.3.12.
Batch writes and schema changes that block or publish an epoch purge the
definition after changing the record, and a definition fill takes its
lease before reading the record, so no fill outlives a newer epoch; on a
cache without generations the fill reads the record again and drops what
it saved when the record moved. A definition holding a blocked collection
is looked at again after a minute.
The epoch record now tells readers the state on its own: an activation
publishes only when no other write is in flight (the last one publishes),
and a published epoch carries the started generation it was published at.
A global collection's definition is one key for every tenant, so each
tenant keeps its field (its own epoch). Inside a transaction that retired
a collection, its documents are read uncached without an epoch read.
`_metadata` is not checked against an epoch any more: batch writes to it
purge each definition they write, purgeCachedCollection('_metadata')
purges every listed definition, and a failed post-commit purge of a
definition is retried once (a definition has no epoch to retire).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The write paths paid the same 6-round-trip collection lookups (J3C-005: update.single 36 → 14, withTransaction(update + get) 26 in the unit model; J3C-007: a sibling read after a write 12). With one round trip per lookup they are back at or below database 7.3.12's counts (bench-core, MariaDB): create 3, update/increase/decrease/delete 4 (7.3.12: 6/4/4/6), withTransaction(update + get of a sibling) 6 (11), ten update + get pairs 80 (110), a sibling read 2 with no adapter read. The unit guards pin the reached figures with assertSame so a regression fails with 7.3.12's figure in the message. The e2e block bounds the same operations on every lane with a document cache, through a counting cache over Redis and the profiler; statements are bounded only on the MariaDB and MySQL lanes, the engines whose 7.3.12 counts were measured, and Mirror is skipped since its destination writes share the counted cache. Co-Authored-By: Claude Opus 5.5 <[email protected]>
…r timeout A write killed between its block and its activation (or whose activation failed, or whose counter was evicted) left #started ahead of #finished for good, and every later write preserved the gap: the collection's document cache stayed off until a flush (B-01, J4L-002; at the start point five reads after three more batch writes and purgeCachedCollection() still reached the database five times). This uses j-db-fix-05's rule for the query cache so both caches behave alike. Tokens record their creation time and a tombstone its block time. A tombstone older than the writer timeout (setCacheWriterTimeout(), default 3600 s, 05's writerTimeout) lapses for readers into an epoch of its own, derived from the tombstone and #finished, so nothing filled before the block is served and any later activation retires what readers filled meanwhile. The next write reconciles: when every other registered write is older than the timeout it releases them and publishes an epoch stamped with the current #started. A write whose registration was released still publishes over an active epoch when it activates, and a missing or unusable record with counters out of step becomes a stamped tombstone so it lapses too. A write younger than the timeout keeps the collection blocked exactly as before; one older is treated as abandoned. Co-Authored-By: Claude Opus 5.5 <[email protected]>
…ery-lib Stage A of the merge-readiness fixes. Parallel relationship population shared authorization, relationship and silence state across coroutines (J4C-006); the branch started at 30f25fd, before the 7.4.0 forward-port, so it lands by merge. RelationshipTests.php: 7.4.0 and this branch each appended tests at the end of the trait; both blocks are kept, with this branch's Swoole imports. Co-Authored-By: Claude Opus 5.5 <[email protected]>
…ry-lib Stage A of the merge-readiness fixes. SQLite contains queries compared JSON text with SQL values and matched no element (J4C-009/J4Q-001), and notContains included NULL arrays on SQLite and MongoDB. Co-Authored-By: Claude Opus 5.5 <[email protected]>
…uery-lib Stage A of the merge-readiness fixes. A nested withTransaction() after a lost connection began a fresh top-level transaction and committed its writes alone while the outer call returned normally (J4L-001). Co-Authored-By: Claude Opus 5.5 <[email protected]>
…y-lib Stage A of the merge-readiness fixes, first link of the cache chain (04 -> 07 -> 09). One cache slot per document, and single-document writes invalidate only their document instead of blocking the whole collection (J3C-004, J3C-006, J3C-008, J4C-004). Co-Authored-By: Claude Opus 5.5 <[email protected]>
…uery-lib Stage A of the merge-readiness fixes, second link of the cache chain. Reads inside a transaction use the cache for documents the transaction has not written (J3C-003), a missing collection costs one _metadata read (JDB13-015), and replica-served reads fill no cache (J4L-006). GeneralTests.php: j-db-fix-03 and this branch each appended a test at the end of the trait; both are kept. Co-Authored-By: Claude Opus 5.5 <[email protected]>
Stage A of the merge-readiness fixes, last link of the cache chain. A collection definition carries its document-cache epoch, so lookups cost one cache round trip again (J3C-001, J3C-002), write and transaction round trips are pinned at 7.3.12's (J3C-005, J3C-007), and an abandoned write lapses after the writer timeout (J4L-002). Co-Authored-By: Claude Opus 5.5 <[email protected]>
Stage A of the merge-readiness fixes. One query-cache hash per
collection scope (J4C-004 query half), purgeCachedCollection() reaches
cached finds (JDB13-016), a query-cache failure falls back to the
database (JDB13-017), and a killed writer's tombstone lapses after the
writer timeout (JDB13-043).
Documents.php conflicts, resolved so every branch's invariant holds:
- purgeCachedCollection(): j-db-fix-09 returns early for _metadata after
purging the definitions and passes the definition key to
advanceDocumentCacheEpoch(); this branch's query-cache invalidation
now runs in both branches, so purgeCachedCollection('_metadata') still
drops cached finds.
- find(): the query-cache fill keeps this branch's fallback try, with
j-db-fix-07's replica guard inside it, so a replica-served result is
never cached and a cache error still only logs a warning.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
…query-lib Stage A of the merge-readiness fixes. document_purge fires after the outermost commit and never after a rollback (JDB13-019), and still fires when the post-commit invalidation fails (JDB13-014). Conflicts, resolved so every branch's invariant holds: - Transactions.php: j-db-fix-07 (transactionWrites) and this branch (documentPurgeEvents) each add a property, one line in the outer-scope open and one entry in both unset() lists; all are kept. - Documents.php deleteDocument(): 7.4.0 returns the deleted document out of the withMutation() callback and fires the delete and related-update hooks after it; this branch queues document_purge inside the callback. The callback now queues the purge when the adapter deleted the row and returns the document, and the old post-callback trigger is dropped, so the purge still fires after the commit and before document_delete. Co-Authored-By: Claude Opus 5.5 <[email protected]>
j-db-fix-01 replaced Database's per-coroutine $silencedListeners array with a State\Value, and j-db-fix-08's queueDocumentPurge() still indexed the property as an array, so every write inside a transaction failed with "Cannot use object of type Utopia\Database\State\Value as array" once both were merged (DocumentPurgeTest: 36 errors, 23 failures). The queued event now captures the silences the writing coroutine sees via silencedListeners()->get(), which is what 08 meant to record. Co-Authored-By: Claude Opus 5.5 <[email protected]>
| $this->assertSame('round '.$round, $database->getDocument('webhooks', 'hook')->getAttribute('name')); | ||
| } | ||
|
|
||
| $this->assertSame(80, $cache->getOperations(), 'Ten updateDocument() + getDocument() pairs (cache.keys_after_1000_writes; 7.3.12: 110 round trips)'); |
There was a problem hiding this comment.
This test requires exactly 80 cache operations for ten update/read pairs. An equivalent cache implementation could add a lease check and fail this assertion while returning the same documents within the stated performance bound. The repository requires tests to check observable behavior, not mirror implementation details or pin version-specific counts. This requirement must be satisfied before merging. The same pattern appears in DocumentCacheInvalidationTest.php’s exact operation counts and DocumentCacheEpochTest.php’s literal blocked: assertions.
Context Used: Call out and harshly judge implementation-coupled tests. We don't mirror source code, configuration, or version pins in assertions. We test observable behavior; use linters for syntax and schema checks. (source)
Prompt To Fix With AI
This is a comment left during a code review.
Path: tests/unit/Documents/DocumentCacheRoundTripTest.php
Line: 217
Comment:
**Tests Pin Cache Operations**
This test requires exactly 80 cache operations for ten update/read pairs. An equivalent cache implementation could add a lease check and fail this assertion while returning the same documents within the stated performance bound. The repository requires tests to check observable behavior, not mirror implementation details or pin version-specific counts. This requirement must be satisfied before merging. The same pattern appears in `DocumentCacheInvalidationTest.php`’s exact operation counts and `DocumentCacheEpochTest.php`’s literal `blocked:` assertions.
**Context Used:** Call out and harshly judge implementation-coupled tests. We don't mirror source code, configuration, or version pins in assertions. We test observable behavior; use linters for syntax and schema checks. ([source](https://app.greptile.com/review/custom-context?memory=instruction-0))
---
For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!
Replaces the string/array database API with
utopia-php/queryvalue objects so adapters, schema, permissions, and queries share one typed surface.Why this approach
Database::VAR_*/INDEX_*/PERMISSION_*and raw attribute arrays could not express the query-lib column, index, and relationship model. Adapters now takeAttribute,Index, andRelationshipobjects; factories (Attribute::string(),Index::key(),Relationship::oneToOne(), …) are the call site. Generic constructors stay for dynamic types.Follow-ups on this branch after the initial migration:
^0.4,^4.0) instead of asterisk (0.4.*)updateDocumentstamps$updatedAtwithDateTime::nowAfter($old)so same-millisecond Redis updates stay uniqueupdateDocumentsstill usesDateTime::now()so bulk patches with unparseable$updatedAtdo not throwValidator\Query\JoinedAttributes)saveWithLeaseleases the document key's own generation, which a rotation never touches, so a schema mutation landing mid-read could negative-cache a collection that by then existedstddev()andvariance()are pinned to the population statistic on every adapter (see below)_metadataentry can no longer outlive its own invalidation under tenant-per-document (see below)Database::createCollection(Collection $collection)only. No string-id overload. AdaptercreateCollection(string $name, array $attributes = [], array $indexes = [])is unchanged._metadatainvalidation under tenant-per-documentUnder shared tables with tenant per document, every collection-metadata cache invalidation was aimed at a key no reader ever used, so anything cached for a collection definition survived its own invalidation for the full 24h TTL.
createDocument()purges throughwithDocumentTenant(), which switches to the document's own tenant. A collection definition is the one rowcreateDocument()accepts without a tenant under tenant-per-document — the check atTraits/Documents.phpexplicitly exemptsself::METADATA— so that switch set the tenant tonullandgetCacheBaseKeys()built the collection key with an empty tenant segment:Two different keys. The epoch the writer rotated and the epoch the readers held were never the same one.
The visible consequences were a stale collection definition after a schema change, and — the one that bit — a negative
$emptymarker. AgetCollection()that legitimately missed once while a project was still provisioning stayed a miss for every later request in that tenant, which surfaces asCollection not foundout of anything that reads the collection.createCollection()performs such a read itself as its pre-existence check, so the collection could be unreadable to its own creator immediately after being created; with the fix reverted,createAttribute()straight aftercreateCollection()throwsNotFound.The fix is to keep the adapter's tenant when the document carries none. A document that does carry a tenant still purges under it, which is what the switch exists for (
GeneralTests::testSharedTablesTenantPerDocumentcovers that across the adapter matrix).This is adapter-independent — it is cache-key arithmetic, not SQL — which is why it reproduced identically on MariaDB, MongoDB and PostgreSQL.
Statistical aggregate contract: population
Bare SQL
STDDEVandVARIANCEare not portable. MySQL and MariaDB read both as the population statistic; PostgreSQL reads both as the sample statistic. Over the same four rows that is 67.0238 against 77.3985 — the same query, two different answers depending on the engine underneath. The query library emits the bare function verbatim (Method::Stddev => 'STDDEV'), so the ambiguity has to be resolved in the adapter.The contract is population.
SQL::applyFindFilters()rewritesMethod::StddevtoMethod::StddevPopandMethod::VariancetoMethod::VarPopbefore the queries reach the builder, so every adapter emitsSTDDEV_POP/VAR_POPexplicitly. Population was chosen because it is what MySQL and MariaDB already returned, what the query library's own ClickHouse builder already emits, and what leaves the existingstddevPop/varPopexpectations unchanged.stddevSamp()andvarSamp()were never ambiguous and are untouched — that is how a caller asks for the sample statistic.Consumers that branch on the adapter to pick an expected value (for instance an
isPostgreSQL()check in a test) can drop the branch and assert the population value on every adapter.Split out of this PR
The typed surface is the change worth reviewing on its own. Three framework layers that were built on top of it were split into separate stacked PRs; all three have since been closed and are not part of the landing plan:
src/Database/ORM,Traits/Entities)src/Database/Repository,src/Database/Seeder)src/Database/Migration,src/Database/Schema,bin/tasks/migrate.php)Nothing in this repository or downstream calls them, so this PR stands alone.
Chain
Landing order, bottom up:
Every consumer above pins
dev-feat-query-libthrough a VCS repository, so this PR has to be tagged before any of them can merge — the monorepo in particular refuses dev-branch pins.utopia-php/database#968 (nested relationship queries) is not stacked on this branch but overlaps it in
Database.phpand the Memory/Mongo/Redis/SQL adapters. Agreed order is #823 first, then #968 rebased on top.Verified
createCollection(CollectionValidationTest, CollectionModelTest, Mirror, CreateCollectionRace)tests/unit/Adapter/PoolTest.php, both seen red before the fix:setTimeouton a pooled handle no longer dials a connection just to answerhasFeature()tests/unit/TenantIdentityTest.phppins PHP's tenant comparison to what_tenant INT UNSIGNEDactually stores, checked against a real MySQLtests/unit/Documents/MetadataTenantInvalidationTest.php— with shared tables and tenant-per-document, a miss cached before provisioning does not outlive the collection being created, and a collection is readable (with the right attribute count) by its own writer after creation and after a schema change. Both cases fail without the fix, the second withNotFound: Collection not foundtests/unit/Documents/NegativeCacheEpochTest.php— a miss observed before a concurrent write rotates the epoch is not negative cached, and a miss with no concurrent write still is. Seen red before the fix (a$emptymarker forwebhooks:hooksurvived the read)tests/unit/StatisticalAggregateContractTest.php— the SQL each of MySQL, MariaDB and Postgres generates forstddev()/variance()is the_POPform, andstddevSamp()/varSamp()are left alone. Seen red before the fix: Postgres emittedSTDDEV("price")(sample) where MySQL emitted the identical text meaning populationAggregationTests::testStddevAndVarianceArePopulationOnEveryAdapter— fixed dataset, asserts the population value and asserts it is not the sample value, across the whole adapter matrixNot verified
Validator\Queries\Documentsand its per-collection cache key — a separate change.Fixes from the 2026-09-23 review
The sections above describe the branch as it was before the review; where they disagree with what follows, what
follows is current.
The review found regressions against 7.x and defects in the features new in 8.0. The fixes ship with regression tests
that fail without them, and the tests that had been weakened or moved to stubbed unit tests run against real engines
again. Upgrading from 7.x: see UPGRADE.md.
Upgrade notes
exception reach the caller and skip the remaining hooks:
index_create,document_read,document_create,documents_create,document_update,documents_update,documents_upsert,document_increase,document_decrease,document_delete,documents_delete,document_find,document_count,document_sum, anddocument_purgefrom a document write orpurgeCachedDocument(). Every other event isolates each hook: its exceptionis swallowed and the remaining hooks still run. An
\Error(aTypeError, for example) always reaches the caller.Differences from 7.x:
\Errorat the isolated events.document_purgefromupdateDocument()anddeleteDocument()fires after the write's transaction has committed.A failing purge listener fails the call, and the write stays committed.
bigint, as in 7.x.ColumnType::BigInteger(value'biginteger') is thein-memory type only. Collection metadata, attribute events and
Attribute::toDocument()carry'bigint', so existingmetadata needs no migration. Write a stored type with
Attribute::persistedType(), and read one withAttribute::normalizeType()/tryNormalizeType(), which accept both spellings.database name (
{cacheName}-cache-{hostname}:{database}:{namespace}:{tenant}:collection:{collection}), and purging adocument advances a per-collection epoch instead of deleting the old key. During a rolling upgrade, a write handled by
a 7.x node purges only the old keys, so an 8.0 node can serve a stale cached document until the cache TTL. After the
last 7.x node has stopped, flush the cache (or deploy without overlap).
Exception\Uniquehas the messageDocument with the requested unique attributes already existson every adapter(was
Unique index violation). Match on the class: catchUniquebeforeDuplicate.Query::DEFAULT_ALIAS, the main collection's alias in generated SQL, istable_main(wasmain).Database::ATTRIBUTE_FILTER_TYPESis renamedATTRIBUTE_FILTER_COLUMN_TYPESand holdsColumnTypecases.new Document()andsetAttribute('$permissions', …)reject non-string permissions withException\Structureinstead of keeping them. Reading stored data never throws:
Document::fromRow()and the newDocument::fromStorage()drop such entries.format(7.x metadata stores'') reads asnull.createCollection()validates attribute types up front, likecreateAttribute().createAttributes()firesattribute_createonce per attribute with aDocumentpayload, then the newattributes_createonce with the list.Hook\Read::applyFilters()takes a requiredPermissionType $forPermission.Validator\Queries\Documentsaccepts joins and aggregations only with its newsupportForJoinsandsupportForAggregationsflags;Databasesets them from the adapter's capabilities.Exception\NotFound(Attribute not found), as on PostgreSQL.Mirrorforwards every configuration setter to its source and destination, and replicatesupsertDocument()andupsertDocumentsWithIncrease().Utopia\Database\PDOis set withPDO::configure().createBuilder()returnsUtopia\Database\Builder\SQLite.Documents
setAttribute(), so awrite to
$permissionsthrough ArrayAccess, a reference,exchangeArray()orunsetleftgetRead()and itssiblings returning the old roles.
getPermissions()returns a de-duplicated list again.Document::fromStorage(), so a stored non-string permission is dropped, as the SQL adapters already did, instead offailing
getDocument(),find(),updateDocument()andupdateDocuments().'unique()'is stored verbatim again, as in 7.x, including for related documents createdthrough relationship attributes. Only an empty id asks the library to generate one.
updateDocuments()with anOperatordecodes the refetched batch once(decode filters ran twice);
count()andsum()on a missing collection throwNotFound; incrementing an unsetoptional number treats it as
0; a case-only$idrename inupdateDocument()is applied instead of dropped; andupdateCollection()validates the metadata it writes.createCollection(),createAttribute()andcreateAttributes()no longer modify the models passed to them.Added filters and adapted index lengths and orders were written back into the caller's objects, so the first create
in a worker rewrote shared config definitions for every later caller.
Exception\Uniquefor unique index violations, like every other adapter.Lifecycle events
testEventscan fail again.triggerHooks()swallowed the assertions the test's recorder made, so the eventcontract passed whatever happened. The test now asserts the whole sequence after the calls.
document_purgefires from document writes again (updateDocument(), for the old and the new id on a rename,updateDocuments(),upsertDocuments(), increments, decrements,deleteDocument()anddeleteDocuments()), onceper purged document after the write's transaction. Cross-region cache relays depend on it.
silent()are back. A hook that also implementsHook\Namedreplaces the hookregistered under its name, in its position, and
silent($callback, ['name'])silences only the named hooks.Caches
find()query cache is scoped like its entries. Epochs and generations were keyed by collection alone, sounder shared tables one tenant's write switched every tenant's cache off and then orphaned their cached results. Every
key now derives from the hostname, database, namespace and tenant.
purgeCachedQueries()also purges thefind()query cache, andlistCollections()is never served stale.TypeRegistry::register()installed every type as aprocess-wide filter, so one handle could replace the built-in
jsonfilter for every other handle in the worker. ADatabasenow resolves filters from its constructor, then itsTypeRegistry, then the global filters; the built-infilter names are reserved; and cache keys follow each type's class. The unused
Embeddablecontract andCustom::columnType()/columnSize()are removed.permission can never publish "not found" for a document another caller may read.
Queries and validation
floatcounts 8 bytes toward the index length; ajsondefault skips the type check only for an array, object orDocument;supportUnsignedBigIntdefaults totrueagain;updateDocuments()anddeleteDocuments()reject joinqueries; and
getDocumentsValidator()takes the joined collections (a list with joins is never cached).having()conditionsfollow the filter rules and compare an aggregate alias of the same query or a
groupByattribute.sum,avg,stddev*andvariance/var*need a numeric attribute that is not an array, andbitAnd/bitOr/bitXoraninteger one; only
countaccepts*.Validator\Querieswith alengthcaps nested groups again.countandcountDistinctreturn0, every other aggregatenull.MariaDB and MySQL returned
18446744073709551615or0for the bitwise aggregates.Database::sum()still returns0for no rows, as in 7.x.Capability::StatisticalAggregatesandCapability::BitwiseAggregates; SQLite reports neither, andfind()rejects those aggregates there. Aggregatealiases are at most 63 characters (PostgreSQL's identifier limit). MariaDB and MySQL errors 1116 and 1191 and
PostgreSQL's DISTINCT/ORDER BY 42P10 (MySQL 3065) map to
Exception\Query, and an unknown column (1054) toException\NotFound.Joins
find()on it would be:every row with a collection-level grant, the rows the caller holds document-level read on with document security, and
Exception\Authorizationotherwise. Adding a join never changes which main rows are visible, infind(),count(),sum()andgetDocument().conditions are, and a right or full outer join checks the main table's and its own conditions in its
ON. Anunreadable document never hides a row an outer join keeps, and a row whose only match is unreadable comes back
unmatched, exactly like a row whose match does not exist.
selected tenant before rows are paired, a later right or full outer join repeats the conditions of the tables joined
before it, and a row without a tenant is never joined. A full outer join combined with a right join reads what a
dedicated database reads, emulated and native.
UNION ALL) return each row once: a later rightjoin's unmatched rows no longer come back twice. Aggregates,
groupBy(),having(),distinct(), ordering and pagingapply once to the whole joined result, as PostgreSQL's native full outer join does. These engines accept one full
outer join per query, and reject a right join after it that joins only on a table cross joined after it and a
distinct()over it ordered by an unselected attribute, withException\Query.$, whichno attribute key can start with; before, an attribute or join alias starting with
foj_ord_read back null.Query::DEFAULT_ALIAS; generated aliases skip declared ones. The tenant and permission conditions name every aliasquoted, so mixed-case and reserved-word aliases work on PostgreSQL and SQLite.
$idand attributes under the alias (alias.$id,alias.attribute), never under a bare name, and joined values are decoded like a direct read (decrypted,JSON and arrays decoded, datetimes formatted, typed). A cursor taken from a joined result pages correctly.
alias.columnis valid exactly whencolumnwould be valid unaliased on the joinedcollection for that query type,
getDocument()validates its join conditions asfind()does, and the numeric andinteger aggregate rules apply to joined attributes. A bare aggregate or
groupByattribute means the maincollection's attribute, else the one joined collection that declares it; an ambiguous or unknown name is rejected.
search()on a joined attribute needs a fulltext index on the joined collection.Shared tables and permissions
upsertDocuments()never registered the tenant write hook,and
Adapter\Poolmade that the normal case, so upserted documents were invisible to document-level readers onMariaDB, MySQL and SQLite.
MySQL, PostgreSQL and SQLite, as 7.x did.
_metadatareads are permission-filtered again, as in 7.x:listCollections()andfind(Database::METADATA)agree with
count(Database::METADATA). Definitions created without a tenant are listed from every tenant of a pool.Database::from()andexecute()are a raw escape hatch that needs authorization disabled (insidegetAuthorization()->skip()). Skipping authorization lifts permissions, never tenant isolation: under shared tablesevery statement from a
from()builder stays in the tenant selected when it was handed out.deleteIndex()do.MongoDB
find(),count()andsum()filter by
$permissionswhether or notHook\Permissionsis registered; writes andgetDocument()are scoped bytenant only, as in 7.x. Writes authorized through update or delete permission no longer silently do nothing when the
caller cannot also read the document.
getSequences()is tenant-safe,listCollections()returns only the definitions the caller may read, and theMongoDB test classes override the trait tests they claim to (the snake_case stubs never did).
SQL adapters
Utopia\Database\PDOreplays session settings madethrough the new
PDO::configure()on the new connection before it retries the call.08; PostgreSQL 57P01–57P05), so MySQL 8.0.24+'s idle disconnect is reconnected without Swoole's library.
setMetadata()reaches the database again. Every prepared statement starts with one/* key: value */comment perentry, ahead of the registered transforms, and comment delimiters, control characters and invalid UTF-8 in keys and
values are neutralised.
_,%and\literally again: the newUtopia\Database\Builder\SQLitedeclaresESCAPE '\'on every LIKE.createAttributes()works for spatial attributes onMariaDB, required spatial attributes on PostgreSQL are nullable columns again, and composite indexes keep the caller's
column order.
Attributes and types
Attribute::TYPES), which every layer now agrees with. Types an adaptercould create but never write or update (timestamp, serial, smallserial) are rejected up front, unknown types fail
with one message everywhere, and
updateAttribute()acceptsidattributes and refuses relationship attributes.tinyInteger(),smallInteger(),decimal(),timestamp(),json(),binary(),enum(),uuid(),uuid7(),serial(),bigSerial(),smallSerial(),array()andtuple(). None shipped in a release.requiredruns before the metadata write, so a failedDROP NOT NULLon PostgreSQL leaves the storeddefinition untouched.
Relationships
getMaxQueryValues()no longer orphans every child. A failed cascade can no longer make a retry skip it.
written. Relinking unchanged children needs only read permission again.
Pools, Mirror and profiling
ReadWritePoolreads your own writes and uses its replicas. Reads stay on the primary for the sticky window aftera write or transaction commits; locking reads and
rawQuery()go to the primary; and metadata and configuration calls(
getHostname(), capability checks, value casting) no longer re-open the window, so replicas are actually read.QueryProfilerkeeps the newest 1000 entries (setCapacity()to change), and pooled connections no longer keepthe profiler of the handle that last borrowed them.
Mirrorinvalidates the query cache it serves, forwards every configuration setter to its source anddestination, opens scoped setters on the source where its writes run, and replicates
upsertDocument()andupsertDocumentsWithIncrease()with their events fired once.Composer
utopia-php/database: 8.*resolves before 8.0.0 is tagged.extra.branch-aliasmapsdev-feat-query-libanddev-mainto8.0.x-dev. The VCSrepositoriesare gone andminimum-stabilityisstableagain; the lock changesonly its provenance. Before the tag, require
8.*withminimum-stability: devandprefer-stable: truein the rootcomposer.json.Tests restored
permissions, object attribute queries and defaults, spatial defaults and index rules, schemaless internal attributes
and TTL indexes, and six relationship tests (array operators on many-to-many and one-to-many, nested permission
enforcement, empty values, no-op nested updates).
testGetDocumentSelect, the PostgreSQL long-name precondition,testTransformations(it now provessetMetadata()reaches the SQL),testCacheFallbackOnFailure(the write pathduring a cache outage),
testLabels,testTrigramIndexValidation,testTTLIndexDuplicatePrevention, and theUniqueassertions oftestDuplicateExceptionMessages,testUniqueIndexDuplicateandtestUniqueIndexDuplicateUpdate.dedicated database and against direct reads, emulated and native full outer joins, and aggregates over a full outer
join with unreadable rows.
Wave 2: fixes and cleanups after the review
Documents and relationships
_metadatawrite runs with validation on again, as in 7.x. The Structure check of the collectiondocument and the query validation of its nested read run in
updateMetadata(),createRelationship()(includingits rollback) and
deleteRelationship(); the library no longer callsskipValidation()internally. A collectiondocument that fails validation fails the schema change: the attribute and index methods and
deleteRelationship()throw
Utopia\Database\Exceptionwith theStructureexception as its previous, andcreateRelationship()throwsFailed to create relationship: ….deleteDocument(), as in 7.x. Thebulk delete it used selected its batch under the caller's read permission, so a related document the caller could
not read was left behind with everything below it and its own
Restrictrelationships were never checked.deleteDocument()loads the document without reading it, checks only the delete permission and cascades belowit, so an unreadable grandchild is deleted with the rest, blocks the delete under
Restrict, or rolls the deleteback when the caller may not delete it.
bigserialis refused whencreateAttribute()/createAttributes()adopt an orphan column. That recovery pathskips the attribute validator and described the column with
getColumnType(), which still mappedbigserial, soover a column that exists in the schema but not in the metadata it was accepted and written to the metadata. It
fails like the other unstorable types, as on 7.x. The remaining handling for the fourteen unstorable types is gone
with it.
MongoDB
upsert filtered by the selected tenant and wrote no
_tenant, so a document of another tenant landed under theselected tenant, or under none. The document's tenant is stamped, falling back to the selected one, and matched,
as
createDocuments()already did and as 7.x did.Aggregations, distinct and search
A select in an aggregation query names only what the query groups by. An aggregation query (an aggregate or a
groupBy()) returns one row per group, holding the groups and the aggregates. Aselect()next to them may nameonly an attribute the query groups by; any other select throws
Utopia\Database\Exception\Query(
Cannot select "<attribute>": an aggregation query can only select the attributes it groups by).*andrelationship wildcards at any depth are accepted and ignored, so appwrite's V20 request filter, which adds them to
every read below response format 1.8.0, keeps working; a join alias's
alias.*is rejected. A bare groupBy namethat only one join declares is that join's column, so
groupBy(['score'])withselect(['note.score'])isgrouped. No select reaches an aggregation's statement. Before, MySQL answered an ungrouped select with a raw 1140,
PostgreSQL with 42803 (a 500 through appwrite), and MariaDB and SQLite returned an arbitrary row's value.
An order in an aggregation query names an aggregate alias or a grouped attribute. Any other order throws
Utopia\Database\Exception\Query(Cannot order by "<attribute>": …) instead of failing in the engine(PostgreSQL 42803, MySQL 1055) or returning an arbitrary order (MariaDB, SQLite).
orderRandom()is accepted.Internal attributes group the joined rows under a join alias.
groupBy(['note.$id']), and the same for$sequence,$createdAt,$updatedAt,$permissionsand, under shared tables,$tenant, return their groupsover every join. They used to fail in the engine (
no such column: note.$id) everywhere but the emulated fullouter join.
A main attribute aggregated under its own name reads the main table.
sum('score', 'score')over a join whosecollection also declares
scorefailed as an ambiguous column. Every bare main-collection aggregate argument isqualified with the main alias when a query joins.
$collectionand$tenantare read only where a table holds them. Aggregates andgroupBy()reject$collection(a read derives it);$tenantin aggregates, groups and joined selects is rejected without sharedtables and works with them.
Join conditions compare columns their tables have, in both the flat and the
on()form: the left columnbelongs to the main collection or to a join declared before, the right column to the joined collection, and a
relationship side without a column is rejected (
Cannot join on virtual relationship attribute: <key>). Theoperator check applies to the flat form too.
Database::sum()validates its attribute as asumaggregate does, and an encrypted attribute of a joinedcollection cannot be filtered under its alias, as on the main collection.
Each aggregate alias names one column of the result. An alias equal to the name a group comes back under, or
to another aggregate's alias, is rejected; the row used to hold only one of the two values or the statement failed.
Bitwise aggregates on PostgreSQL name their input count
$inputs:<n>, which no engine truncates.Aggregation and distinct reads next to a
search()or a vector query run on every engine. The relevanceprojection and order, and the vector distance order, were added to every read without an explicit order, so an
aggregate next to a search failed with a raw 1064 (MariaDB, MySQL) or 42803 (PostgreSQL), a
groupBy()returnedan extra
_relevancecolumn, adistinct()read returned each selection once per relevance value, and a distinctread next to a vector query failed on PostgreSQL (42P10). An aggregation query returns its groups and aggregates
only; a distinct read returns each selection once, ordered by its explicit orders, with the ordinary cursor.
Vector distance orders row reads only.
A
search()read paged with a cursor lists every match once, and a search only filters, as in 7.x. Without anexplicit order the read was ordered by relevance while the cursor compared
$sequence, so pages skipped andrepeated rows. No read is ordered by relevance by itself, rows carry no
_relevance, and a search read is orderedby its explicit orders and then
$sequence.getSearchRelevanceRaw()(new in the PR) andPostgres::getFulltextValue()are removed; the query builders normalize search terms.A
distinct()read whose select names only joined columns returns each of them once, under its alias, andaccepts a joined internal attribute (
alias.$id). The projection replaced the select, but the caller's select wasforwarded as well and compiled raw, which duplicated the columns and put a joined value on the main document's
bare key.
A filter on a joined column is checked against the joined collection's attribute. The joined-column branch of
the filter validator checked only the value count, so
equal('book.pages', ['abc'])reached PostgreSQL as 22P02,a bad date as 22007,
startsWithon an integer as 42883 (all 500 through appwrite), and a nested list asUnknown PDO Type for arrayon every engine. The validator now applies the main collection's type, size, array andcontainsrules to the joined attribute, and rejects vector queries on joined attributes, which no adapter compiles.Query cache
find()caches its results per tenant, but a write invalidated only the scope of the tenant selected on the writing
Database, so a written tenant's cached results stayed in place until the region TTL: new and changed rows stayedinvisible to its readers, and a reader whose grant an upsert had just revoked was still served the document. The
tombstones and epochs now cover each written document's collection in the scope of its tenant (its own, else the
selected one), and every other tenant's cached results stay in place.
options.relatedCollectiononly for attribute events, so a document with an attribute calledoptionsrefreshesits own collection and nothing else.
Removed code
getSQLConditionsForCollection()and everything behind it, about 800lines across the SQL adapters),
Utopia\Database\Loading\*andUtopia\Database\Traits\Asyncare removed.Nothing wired them in.
utopia-php/asyncstays a dependency ofMirrorand the relationship hook.Tests
convert*ToExceptionsattributes and the PHPUnit 12 configuration had no replacement, so such issues were countedbut never failed a run.
phpunit.xmlnow setsfailOnWarning,failOnNotice,failOnDeprecation,failOnRisky,failOnPhpunitNoticeandfailOnPhpunitDeprecation, scoped tosrc/andtests/; paratest computes the CIlanes' exit code from the same settings. Fixed at the source while turning this on:
SpatialFilterTest'sexpectation-less mock,
PoolTest'ssetAccessible(), 17 e2e calls of the deprecatedQuery::contains(), and analias-length test on the lanes without aggregations.
Documentation
UPGRADE.mdwalks a 7.x consumer through every change with the two forms side by side,CHANGELOG.mdlists therelease with the fixes to the 8.0 pre-releases in their own subsection, and
README.mddescribes the 8.0 API.🤖 Generated with Claude Code