You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Async-durability inserts 5-8% slower since the readiness merges (not bisected) #374
Found by the readiness benchmark checkpoint on the benchmark host (2026-10-05, report in the business repo docs/performance). insert_single p50 0.224 -> 0.236 ms (ratio 1.053, 95% CI 1.009..1.106), 4 writers 1.076. Synchronous durable inserts unchanged. Bisect across the 24 merges between 25058f1 and 57ad2e9 on the benchmark host and fix or justify.
Bisected on the benchmark host and accepted as the cost of a deliberate change. The slowdown is 1.3-2% on async-durability inserts, not 5-8% (current main vs the range start: 1.015, 95% interval 1.009..1.022); synchronous durable inserts are unchanged. It comes from #275, the per-query and server memory limits, which meter every allocation (about 3 microseconds per insert; one thread-local add per allocation is the floor). Closing as justified: it buys the memory budgets and resource_exhausted refusals.
Activity
jsam commentedon Oct 6, 2026
Bisected on the benchmark host and accepted as the cost of a deliberate change. The slowdown is 1.3-2% on async-durability inserts, not 5-8% (current main vs the range start: 1.015, 95% interval 1.009..1.022); synchronous durable inserts are unchanged. It comes from #275, the per-query and server memory limits, which meter every allocation (about 3 microseconds per insert; one thread-local add per allocation is the floor). Closing as justified: it buys the memory budgets and resource_exhausted refusals.