logger use dict for loggers instead of WeakValueDictionary

[socketpair]

BPO	11950
Nosy	@terryjreedy, @vsajip, @socketpair

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = 'https://github.com/vsajip'
closed_at = <Date 2011-04-29.16:40:37.106>
created_at = <Date 2011-04-28.19:29:16.288>
labels = ['invalid', 'type-feature', 'library']
title = 'logger use dict for loggers instead of WeakValueDictionary'
updated_at = <Date 2011-04-29.23:38:58.577>
user = 'https://github.com/socketpair'

bugs.python.org fields:

activity = <Date 2011-04-29.23:38:58.577>
actor = 'terry.reedy'
assignee = 'vinay.sajip'
closed = True
closed_date = <Date 2011-04-29.16:40:37.106>
closer = 'vinay.sajip'
components = ['Library (Lib)']
creation = <Date 2011-04-28.19:29:16.288>
creator = 'socketpair'
dependencies = []
files = []
hgrepos = []
issue_num = 11950
keywords = []
message_count = 4.0
messages = ['134714', '134788', '134807', '134823']
nosy_count = 3.0
nosy_names = ['terry.reedy', 'vinay.sajip', 'socketpair']
pr_nums = []
priority = 'normal'
resolution = 'not a bug'
stage = None
status = 'closed'
superseder = None
type = 'enhancement'
url = 'https://bugs.python.org/issue11950'
versions = ['Python 3.3']

[socketpair]

logger use dict for loggers instead of WeakValueDictionary

Now, when object returned by getLogger() gets unreferenced - it will not be garbage collected, as reference is stored in logging module. This will results in leaked file object, that cannot be closed in easy way. Some modules (notable yum) opens dozen of loggers that prevents me to unmount directory where logs placed. Now, I use workaround:

-----------------------
try:
# work with yum
finally:
yum_base_cli.close()
yum_base_cli.closeRpmDB()
yum_base_cli.doUnlock()
del yum_base_cli
# gc.collect() may be called here, but it does not actually change anything....
for logger_name,logger in cli.logging.Logger.manager.loggerDict.iteritems():
if not logger_name.startswith('yum'):
continue
for h in logger.handlers:
h.flush()
h.close()
-----------------------
The problem in that logging module stores many references in lists and dicts. So just replacing {} with WeakValueDictionary()
does not eliminate problem fully.

I have checked, problem exists in all python versions

[vsajip]

I had to delete my previous response to the initial post, as something got mangled in it. Here's what I meant to say:
----------------------------------------------
It sounds to me as if there's a problem in yum that needs sorting out: I'm not sure why it would need dozens of loggers.

As your workaround snippet shows, leaking files is not due to loggers but due to handlers - they're a different animal. Handlers need not be leaked - they can just be closed when you're done with them. Handlers *are* stored with weak references.

The design contract for loggers precludes using weak references for them. For example, if I use a library and from my logging initialization code get a reference to its top-level logger and e.g. set that logger's level, then when I've finished, there may not be any other references to that logger until I do an import of some library module. I wouldn't want the reference to the logger to go away, because then the library wouldn't know my intention regarding the level setting.

Normal practice would be to add a few handlers to yum's top-level logger and let descendant loggers inherit them. I'm not familiar with yum, so I don't know if/why it deviates from this common and best practice approach.

As you state, replacing {} with WeakValueDictionary() wouldn't solve the problem, and would violate the design assumptions/contract relating to loggers.

By the way, the logging module does not store loggers in lists and dicts - just in a single dict in a manager. Handlers are stored in weak reference dicts as well as lists but that does not mean they are guaranteed to be garbage collected, since there could also be references to them in user code. Likewise for loggers - typically there would be references to them in lines like

logger = logging.getLogger(__name__)

in numerous modules. Even if yum consists of dozens of modules (which perhaps explains dozens of loggers), they needn't leak files, as just adding a single FileHandler to yum's top-level logger would serve to capture logging events from the entire yum package hierarchy.

[socketpair]

I'm not YUM developer, I'm very sad user of YUM API. Also, I'm novice in python logging complex system.

1. There is no way to remove logger once it added via getLogger(). why?
2. When yum should close handlers? In destructor (cleanup function) of module?
3. Why this code does not close log files?
   ----------------

def test():
    import yum
    # work with yum

test()
gc.collect() # to sure that nothing have extra references.
# yum logs still alive in spite of that yum module (and imported logging module) 'is unloaded'

---

3. Also, logger.handlers is not documented...it's definitely the bug as I think.

[terryjreedy]

Mark, this is a feature request to change the storage of loggers. A behavior issue (bug report) must report a discrepancy between doc and behavior. Vinay rejected that request as not really possible.

Questions about using yum should go to a yum list or to python-list.
The tracker is not a tutorial list.

"logger.handlers is not documented"
Please look at section 15.9. logging.handlers — Logging handlers
Are you referring specifically to how they are stored?
We do not usually document the internal details of modules.