TimedRotatingFileHandler rotating on use not time · Issue #84649 · python/cpython · GitHub

BPO	40469
Nosy	@vsajip, @stevendaprano, @martonivan
PRs	gh-84649: Make TimedRotatingFileHandler use CTIME instead of MTIME #24660
Files	logging_config.ymal app.py: Eg code to use the ymal file config.json: JSON version of test config log_40469.py: Modified test script log_40469_new_instance.py log_40469_single.py

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 = None
closed_at = None
created_at = <Date 2020-05-01.15:02:02.675>
labels = ['3.8']
title = 'TimedRotatingFileHandler rotating on use not time'
updated_at = <Date 2022-03-31.20:48:10.169>
user = 'https://bugs.python.org/markhallett'

bugs.python.org fields:

activity = <Date 2022-03-31.20:48:10.169>
actor = 'steven.daprano'
assignee = 'none'
closed = False
closed_date = None
closer = None
components = []
creation = <Date 2020-05-01.15:02:02.675>
creator = 'markhallett'
dependencies = []
files = ['49111', '49112', '50196', '50197', '50208', '50209']
hgrepos = []
issue_num = 40469
keywords = []
message_count = 11.0
messages = ['367849', '367922', '398684', '399306', '399481', '399496', '399499', '399500', '399506', '399531', '416458']
nosy_count = 4.0
nosy_names = ['vinay.sajip', 'steven.daprano', 'markhallett', 'martonivan']
pr_nums = ['24660']
priority = 'normal'
resolution = None
stage = 'patch review'
status = 'open'
superseder = None
type = None
url = 'https://bugs.python.org/issue40469'
versions = ['Python 3.8']

Linked PRs

Using the attached confing ymal file, the log file only rolls if the log file is not written to for over a minute, logging regulary (eg every 30s) prevents the correct rolling of the log.

Eg code to run is now attached

With this slightly modified test script and the JSON configuration based on the YAML, I got the following results after running for a while:

$ more info.log*
::::::::::::::
info.log
::::::::::::::
2021-08-01 07:13:45 INFO Message no. 9
2021-08-01 07:14:15 INFO Message no. 10
::::::::::::::
info.log.2021-08-01_07-09
::::::::::::::
2021-08-01 07:09:45 INFO Message no. 1
2021-08-01 07:10:15 INFO Message no. 2
::::::::::::::
info.log.2021-08-01_07-10
::::::::::::::
2021-08-01 07:10:45 INFO Message no. 3
2021-08-01 07:11:15 INFO Message no. 4
::::::::::::::
info.log.2021-08-01_07-11
::::::::::::::
2021-08-01 07:11:45 INFO Message no. 5
2021-08-01 07:12:15 INFO Message no. 6
::::::::::::::
info.log.2021-08-01_07-12
::::::::::::::
2021-08-01 07:12:45 INFO Message no. 7
2021-08-01 07:13:15 INFO Message no. 8

The test script logs every 30 seconds, so I don't see what the problem is. Can you describe it more clearly?

The assumed behaviour of TimedRotatingFileHandler is to rotate log files older than configured. Even when the script is executed multiple times.

self.rolloverAt (the variable that defines when the rollover should be performed) is set after each rollover and right after initializing the file handler instance.

If the instance is initialized once (like in your script) and a rollover is performed without having the object destroyed, the logger works fine, the rollover is done and the next round is scheduled properly.

The case is not so lucky if the script runs multiple time or the logger itself is initialized multiple times during one execution. In this case, since the MTIME is read each time when the TimedRotatingFileHandler's init is called, and the file is modified (by having a new line added to the end of the file). The next execution will read the new MTIME and will never perform any rollover.

I've slightly modified your example script to demonstrate these use-cases.

Example 1: Log a single line with the script, but execute it multiple times with delays between the execution!
In bash: for i in {1..101}; do python log_40469_single.py $i; sleep 30; done

Example 2: Log multiple lines, but reinitiate the logger object between the events!
See log_40469_new_instance.py

The assumed behaviour of TimedRotatingFileHandler is to rotate log files older than configured. Even when the script is executed multiple times.

Ah, but is it? The purpose of TimedRotatingFileHandler is to rotate files based on time intervals, and is normally assumed to be for long-running processes (where the rotation allows one to focus on logs output during a period of interest). What's the practical, real-world use-case for the example scripts you added?

Well, there is nothing in the documentation that would say the logrotation can be only performed by continously running scripts, thats why. :)

There are scripts or even daemons written in Python that need to be restarted from time to time. If the logrotate period is greater than restart period the log files will never be rotated, however they could be.

We were looking for something that works like the logrotate do, but without involving any external logic, but the one written in python. And TimedLogrotateHandler appears to be one of those, unless it checks the MTIME of the files instead of the CTIME.

What is the logic behind checking the MTIME instead of CTIME when one is calculating the age of a file when tries to calculate whether it has to be rotated or not? If the only purpose of this initial rollover calculation would be to start a predefined long period (as you suggest), now() (time of execution) should used instead and no file attribute should be involved. From the code I have to assume that the original author also tried to prepare for the cases where the execution is interrupted and/or restarted and (s)he wanted the TimedLogrotateHandler to be resumed where it was during the previous execution. Dnn't you agree?

Well, there is nothing in the documentation that would say the logrotation can be only performed by continously running scripts, thats why. :)

Well, what's the need to rotate based on time when you just run scripts sporadically or they're very short-lived? Not trying to argue, just trying to understand.

There are scripts or even daemons written in Python that need to be restarted from time to time. If the logrotate period is greater than restart period the log files will never be rotated, however they could be.

Yes, but generally daemons are long-lived relative to the rotation interval, aren't they? "From time to time" certainly implies that to me.

What is the logic behind checking the MTIME instead of CTIME when one is calculating the age of a file when tries to calculate whether it has to be rotated or not?

I can't remember, that code is over 12 years old and that specific decision wasn't documented :-(

From the code I have to assume that the original author also tried to prepare for the cases where the execution is interrupted and/or restarted and (s)he wanted the TimedLogrotateHandler to be resumed where it was during the previous execution. Dnn't you agree?

I take it you mean TimedRotatingFileHandler. Certainly it assumes that the script might be restarted, which is why it opens the file in append mode.

If the computation using MTIME were changed to use CTIME, that would be a behaviour change, which could conceivably break someone's code that relied on current behaviour (unlikely, but you just never know). I've no other objection to using CTIME rather than MTIME - perhaps I'm just being hyper-cautious about breakage?

I just checked - the relevant code (using MTIME) was added in 2010, as a fix for bpo-8117 - around the time that Python 2.6 was released.

Well, what's the need to rotate based on time when you just run scripts sporadically or they're very short-lived? Not trying to argue, just trying to understand.
Yes, but generally daemons are long-lived relative to the rotation interval, aren't they? "From time to time" certainly implies that to me.

In one of our use-cases, where I faced with this bug/behaviour of TimedRotatingFileHandler, we are using it in a daemon for daily logrotate. The aim is to have the logs separated day by day, just like the majority of our other components running on a linux. We've just realized that some days are missing from the series of logs and the investigation led to the service restart and the MTIME checking that we are talking about here. When a logger is configured to rotate daily one just simply assume that it will rotate daily regardless the daemon restart. (As en extreme(?) example, when a the system - and the daemon of course - is restarted each day - or just never running for a whole day - and btw it logs often, the log file won't be ever rotated and can inrease over every limits.)

I've quickly googled around and found someone who execute some logic from cron and expects the loghandler to rotate his files, that wouldn't happened for the same reason:
https://stackoverflow.com/questions/30569164/logging-handler-timedrotatingfilehandler-never-rotates-the-log

I just checked - the relevant code (using MTIME) was added in 2010, as a fix for bpo-8117 - around the time that Python 2.6 was released.
As I understand bpo-8117, the goal was to do the rollover even after script restarts. (To be frank, only an extreme case was mentioned in the issue opener entry, when the rollover period had been reached when the script wasn't running.) I believe for this purpose MTIME wasn't the best choice, but using CTIME instead would make the script do what is expected to.

I take it you mean TimedRotatingFileHandler. Certainly it assumes that the script might be restarted, which is why it opens the file in append mode.

If the computation using MTIME were changed to use CTIME, that would be a behaviour change, which could conceivably break someone's code that relied on current behaviour (unlikely, but you just never know). I've no other objection to using CTIME rather than MTIME - perhaps I'm just being hyper-cautious about breakage?

You are far better and more experienced in maintaining widely used opensource codebases so I wouldn't argue with you, but... :)

In case of scripts that create the log file for themselves the change would have no effect.
In case of scripts that are using already existing files for logging (by opening them for append) could change the behaviour, but just slighly. The first rollover would happen before adding the first new log line if the file age reached the configured time, and not when the age of last log line written to the file before having the script started reached the same value. No further rotation would be effected.

I _believe_ no one expects the current initial behaviour in case of resumed log files, and for sure, the documentation doesn't cover this part so I don't know how to proceed.

I _believe_ (again) that changing the behaviour wouldn't break the script outside, but rather would help them working the way how the developers originally assumed they did.

You are far better and more experienced in maintaining widely used opensource codebases so I wouldn't argue with you, but... :)

Well, if a change introduces a problem, it will be my lookout, won't it? Just as the original patch for bpo-8117 is proving to be :-)

OK, how about this? We update the PR to add a new use_ctime=False keyword argument to TimedRotatingHandler.__init__(). If nothing else changes, the behaviour is as it is currently. If use_ctime=True is passed in the instantiation, then the constructor will use CTIME rather than MTIME. Will that work for you? I'd be OK with that approach.

See this thread on Discuss:

https://discuss.python.org/t/logging-timedrotatingfilehandler-never-rotates-in-certain-cases/14747/1

pythongh-84649: Make TimedRotatingFileHandler use CTIME instead of MT…

pythongh-84649: Fix unstable test_rollover_based_on_st_birthtime_only (…

pythongh-84649: Use statx() in TimedRotatingFileHandler if available …

gh-84649: Fix unstable test_rollover_at_midnight (GH-151813)

The test is still flaky in the CI, for example this run from today (on main after I merged a PR):

======================================================================
FAIL: test_rollover_at_midnight (test.test_logging.TimedRotatingFileHandlerTest.test_rollover_at_midnight)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "C:\a\cpython\cpython\Lib\test\test_logging.py", line 6749, in test_rollover_at_midnight
    self.assertIn(f'testing2 {i}', line)
    ~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^
AssertionError: 'testing2 0' not found in '2026-06-29 09:32:51,103 testing1 0\n'

----------------------------------------------------------------------

Another in #152760 / https://github.com/python/cpython/actions/runs/28508559349/job/84502982384?pr=152760:

======================================================================
FAIL: test_rollover_at_midnight (test.test_logging.TimedRotatingFileHandlerTest.test_rollover_at_midnight)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "D:\a\cpython\cpython\Lib\test\test_logging.py", line 6749, in test_rollover_at_midnight
    self.assertIn(f'testing2 {i}', line)
    ~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^
AssertionError: 'testing2 0' not found in '2026-07-01 10:00:19,101 testing1 0\n'

pythongh-84649: Fix unstable test_rollover_at_midnight (pythonGH-151813)

Windows / Build and test (Win32, switch-case) on the main branch: https://github.com/python/cpython/actions/runs/29167896618/job/86584142360?pr=153591

test_logging fails with:

FAIL: test_rollover_at_midnight (test.test_logging.TimedRotatingFileHandlerTest.test_rollover_at_midnight)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "D:\a\cpython\cpython\Lib\test\test_logging.py", line 6749, in test_rollover_at_midnight
    self.assertIn(f'testing2 {i}', line)
    ~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^
AssertionError: 'testing2 0' not found in '2026-07-11 21:19:17,102 testing1 0\n'

cc @serhiy-storchaka

I know. It is annoying.

My previous fix was not enough. Claude initially proposed a fix that deferred the purpose of the test, so I need to be very cautious here. I'll fix it when my current blackout ends and I could test it locally on Windows.

gh-84649: Fix unstable test_rollover_at_midnight (GH-154463)

The root cause was that on Windows/NTFS, if delete and recreate a file, it keeps the birth time of the old file. 🧟

TimedRotatingFileHandler rotating on use not time · Issue #84649 · python/cpython · GitHub

Repository navigation

TimedRotatingFileHandler rotating on use not time #84649

Linked PRs

21 remaining items

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

TimedRotatingFileHandler rotating on use not time · Issue #84649 · python/cpython · GitHub

Repository navigation

TimedRotatingFileHandler rotating on use not time #84649

Description

Linked PRs

Activity

markhallett commented on May 1, 2020

markhallett commented on May 2, 2020

vsajip commented on Aug 1, 2021

martonivan commented on Aug 10, 2021

vsajip commented on Aug 12, 2021

martonivan commented on Aug 13, 2021

vsajip commented on Aug 13, 2021

vsajip commented on Aug 13, 2021

martonivan commented on Aug 13, 2021

vsajip commented on Aug 13, 2021

stevendaprano commented on Mar 31, 2022

21 remaining items

StanFromIreland commented on Jun 29, 2026

hugovk commented on Jul 1, 2026

vstinner commented on Jul 11, 2026

serhiy-storchaka commented on Jul 12, 2026

serhiy-storchaka commented on Jul 22, 2026

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions