openml registers handler on root logger #736

Activity

1. mfeurer commented on Jul 12, 2019

   

   mfeurer

   on Jul 12, 2019

   Collaborator

   More actions

   Thanks for bringing that up. Actually, the package should have a good default logging behavior, but should allow flexible configuration and fully disabling it, too. I currently don't know the best way to achieve this unfortunately.
2. PGijsbers commented on Jul 12, 2019

   

   PGijsbers

   on Jul 12, 2019

   CollaboratorAuthor

   More actions

   What do you consider 'good default logging behavior'? For me it would be:

   • Have a package's root logger (e.g. logging.getLogger('openml'), not logging.getLogger()) configured at logging.DEBUG. This means any logging.Handler registered will by default pick up all messages, limited by the logging.Handler.level.
   • Have each module log to its own logger (i.e. logging.getLogger(__name__)).
     As the name will be openml. ... any logging.Handler registered with the package root logger will also pick up on messages sent by the package root logger (it keeps a hierarchy based on name). But if we're alternatively only interested in just registering for messages from submodules, we can directly register a logging.Handler with their logger.
   • My preference would be that any logging.Handler registers only with the package root logger, and only if the openml config file specifies it (but this can be a default). By default, I would only consider using a logging.FileHandler to write the log to a file. Perhaps writing to a new file each day, and removing any files older than a week automatically (again exposing freedom for how long (if at all) to retain logs through the configuration file).
3. mfeurer commented on Jul 15, 2019

   

   mfeurer

   on Jul 15, 2019

   Collaborator

   More actions

   Have a package's root logger (e.g. logging.getLogger('openml'), not logging.getLogger()) configured at logging.DEBUG. This means any logging.Handler registered will by default pick up all messages, limited by the logging.Handler.level

   I would not show debug messages by default, but I agree on the logger name.

   Have each module log to its own logger (i.e. logging.getLogger(name)).
   As the name will be openml. ... any logging.Handler registered with the package root logger will also pick up on messages sent by the package root logger (it keeps a hierarchy based on name). But if we're alternatively only interested in just registering for messages from submodules, we can directly register a logging.Handler with their logger.

   Agreed.

   My preference would be that any logging.Handler registers only with the package root logger, and only if the openml config file specifies it (but this can be a default). By default, I would only consider using a logging.FileHandler to write the log to a file. Perhaps writing to a new file each day, and removing any files older than a week automatically (again exposing freedom for how long (if at all) to retain logs through the configuration file).

   I didn't think of logging to the openml directory by default, but that appears to be a great idea. The user could still configure the package to output to the command line.
4. PGijsbers commented on Jul 17, 2019

   

   PGijsbers

   on Jul 17, 2019

   CollaboratorAuthor

   More actions

   I would not show debug messages by default, but I agree on the logger name.

   I'm not advocating showing (to console?) debug messages by default either. There's two levels to set, that of the logger and those of the logger.Handler. No messages would be shown if no handlers are set to the logger. A message is only processed if the level is met by both the log-object and the handler. By default no handlers would be set. To illustrate:

   >>> import logging
   >>> openml_log = logging.getLogger('openml')
   >>> openml_log.setLevel(logging.DEBUG)
   >>> openml_log.error("With the exception of error-level messages")
   With the exception of error-level messages
   >>> openml_log.info("No messages are written to console.")
   >>> # if you want to change this default behavior, and have not even error messages
   ... # written to stdout, you can add a `logging.NullHandler()`.
   ... # Now we add a 'logging.StreamHandler' to actually write to stdout
   ...
   >>> import sys
   >>> sh = logging.StreamHandler(sys.stdout)
   >>> sh.setLevel(logging.INFO)
   >>> openml_log.addHandler(sh)
   >>> openml_log.info("Now info messages are also written to stdout.")
   Now info messages are also written to stdout.
   >>> openml_log.debug("but not debug messages.")  # because `sh.level == logging.INFO`
   >>> sh.setLevel(logging.DEBUG)
   >>> openml_log.debug("Now even DEBUG messages are written to stdout.")
   Now even DEBUG messages are written to stdout.
   >>> openml_log.setLevel(logging.ERROR)
   >>> openml_log.warning("By setting the log level of the log differently, handlers wont receive all messages.")  # because `openml_log.level == logging.ERROR`
   >>> # Nothing written to stdout

   By setting the log level of the log (not the handler!) to logging.DEBUG, we allow any handler to subscribe to any level of messages it is interested in.

   ---

   I think the proposed changes will be relatively easy to implement. I don't have the time currently (when I do, I'll first finish the 'short names' PR). It's a low priority issue but if it stays open I can tackle it end of the month/in August.
5. mfeurer commented on Jul 17, 2019

   

   mfeurer

   on Jul 17, 2019

   Collaborator

   More actions

   Got it, that totally makes sense.
6. PGijsbers

   closed this as completedin #861on Nov 5, 2019