[tor-bugs] #20540 [Metrics]: define log-levels for all java metrics-products

[Tor Bug Tracker & Wiki]

#20540: define log-levels for all java metrics-products
-------------------------+-----------------------------------
 Reporter:  iwakeh       |          Owner:
     Type:  enhancement  |         Status:  needs_information
 Priority:  Medium       |      Milestone:
Component:  Metrics      |        Version:
 Severity:  Normal       |     Resolution:
 Keywords:               |  Actual Points:
Parent ID:               |         Points:
 Reviewer:               |        Sponsor:
-------------------------+-----------------------------------

Comment (by iwakeh):

 Replying to [comment:3 karsten]:
 > Replying to [comment:1 iwakeh]:
 > ... Here's a suggestion:
 >
 >  - trace: I suggest disregarding this level, because the slf4j
 developers themselves only added it to "bow to popular demand".  It seems
 we can easily get away by just using debug instead.

 Agreed.

 >  - debug: Let's use this for detailed messages to debug a problem, under
 the assumption that these logs are typically disabled in production and
 only enabled when debugging a problem.

 Agreed.

 >  - info: It seems useful to log whenever a process or major step starts
 or ends, but under the assumption that these logs will only be written to
 files and not mailed out to operators.  It might be a good requirement to
 write info-level logs in a way that operators can understand them without
 having to read any code.

 Fine.

 > One example where info might be too high is where metrics-lib informs us
 which `DescriptorCollector` implementation it's serving us, because that's
 something the operator wouldn't normally care about.

 The logging statement you refer to can be useful to the operator once
 there are other metrics-lib implementations available and the operator
 choses to use metrics-lib-ng instead of the regular implementation.  Or,
 in case we provide other parsing implementations, too.

 (In order to make it really code independent metrics-lib would need to
 supply an identifier, but when different implementations are used the full
 class-name can just be stated in the documentation to that implementation
 and operators wouldn't need to know that this is a java class name.)

 >  - warn: We could use warn to inform the operator of a problem that we
 were able to recover from but that they should be aware of.  The warning
 should be written in a way that the operator understands, and it should be
 something that the operator can do something against.  Stated differently,
 we should expect to be contacted by operators who are unclear what to do
 about a given warning.  And if they cannot do anything against it, maybe
 it should be an info message rather than a warning.  We might want to
 recommend that operators include warnings in any automated notifications
 they receive from their service instance.

 Agreed. Even require that the 'cure' is added to the warning, e.g.
 "provide more disk space" etc.  Requiring this will give a good
 distinction between warnings and info messages.

 >  - error: We should use error for problems that we cannot recover from.
 Otherwise they're similar to warnings.

 Maybe, define errors for the unusual (errors would be logged in some
 places where we do not really expect anything) and ask operators to post
 these (ml or trac)?

 >
 > > * maybe, log statistics to separate log files (as in Onionoo).
 >
 > Are there log domains of some sort?  It seems that we should leave the
 decision of log files to the operator, but could say that these log
 messages go into a "statistics" log domain that the operator may log to
 the same or a different file.  By the way, we'd probably want to log these
 messages on info level.

 Yes, you can call it domains.  There should be domains for certain log
 lines (like statistics) that can be easily directed to different log-line
 consumers (other files, databases, servers, etc.).

 >
 > > * what else?
 >
 > Maybe one thing:
 >
 >  * Log levels used by metrics-lib, where a problem with parsing a
 descriptor can have different consequences depending on the application.
 In other words, if we log a warning or even error, we should give the
 application the opportunity to tone down that warning, or ignore it,
 because it doesn't care as much.  What we could do is use a log domain
 "parsing", or we could let applications define logging by logger name and
 tone down `org.torproject.descriptor.*` loggers.

 Maybe, for metrics-lib we should not log too much but rather throw
 exceptions.  That would be the right way to make the application to deal
 with problems.
 But this is "error handling" rather than logging.

--
Ticket URL: <https://trac.torproject.org/projects/tor/ticket/20540#comment:4>
Tor Bug Tracker & Wiki <https://trac.torproject.org/>
The Tor Project: anonymity online