So merging the study from @bmw and me, here is what happened.
Each invocation of `certbot.logger.post_arg_parse_setup` create a file handler on `letsencrypt.log`. This function also set an atexit handler invoking `logger.shutdown()`, that have the effect to close all logger file handler not already closed at this point. This method is supposed to be called when a python process is close to exit, because it makes all logger unable to write new logs on any handler.
Before #6667 and this PR, for tests, the atexit handle would be triggered only at the end of the pytest process. It means that each test that launches `certbot.logger.post_arg_parse_setup` add a new file handler. These tests were typically connecting the file handler on a `letsencrypt.log` located in a temporary directory, and this directory and content was wipped out at each test tearDown. As a consequence, the file handles, not cleared from the logger, were accumulating in the logger, with all of them connected to a deleted file log, except the last one that was just created by the current test. Considering the number of tests concerned, there were ~300 file handler at the end of pytest execution.
One can see that, on prior #6667, by calling `print(logger.getLogger().handlers` on the `tearDown` of these tests, and see the array growing at each test execution.
Even if this represent a memory leak, this situation was not really a problem on Linux: because a file can be deleted before it is closed, it was only meaning that a given invocation of `logger.debug` for instance, during the tests, was written in 300 log files. The overhead is negligeable. On Windows however, the file handlers were failing because you cannot delete a file before it is closed.
It was one of the reason for #6667, that added a call to `logging.shutdown()` at each test tearDown, with the consequence to close all file handlers. At this point, Linux is not happy anymore. Any call to `logger.warn` will generate an error for each closed file handler. As a file handler is added for each test, the number of errors grows on each test, following an arithmetical suite divergence.
On `test_sdists.py`, that is using the bare setuptools test suite without output capturing, we can see the damages. The total output takes 216000 lines, and 23000 errors are generated. A decent machine can support this load, but a not a small AWS instance, that is crashing during the execution. Even with pytest, the captured output and the memory leak become so large that segfaults are generated.
On the current PR, the problem is solved, by resetting the file handlers array on the logging system on each test tearDown. So each fileHandler is properly closed, and removed from the stack. They do not participate anymore in the logging system, and can be garbage collected. Then we stay on always one file handler opened at any time, and tests can succeed on AWS instances.
For the record, here is all the places where the logging system is called and fail if there is still file handlers closed but not cleaned (extracted from the original huge output before correction):
```
Logged from file account.py, line 116
Logged from file account.py, line 178
Logged from file client.py, line 166
Logged from file client.py, line 295
Logged from file client.py, line 415
Logged from file client.py, line 422
Logged from file client.py, line 480
Logged from file client.py, line 503
Logged from file client.py, line 540
Logged from file client.py, line 601
Logged from file client.py, line 622
Logged from file client.py, line 750
Logged from file cli.py, line 220
Logged from file cli.py, line 226
Logged from file crypto_util.py, line 101
Logged from file crypto_util.py, line 127
Logged from file crypto_util.py, line 147
Logged from file crypto_util.py, line 261
Logged from file crypto_util.py, line 283
Logged from file crypto_util.py, line 307
Logged from file crypto_util.py, line 336
Logged from file disco.py, line 116
Logged from file disco.py, line 124
Logged from file disco.py, line 134
Logged from file disco.py, line 138
Logged from file disco.py, line 141
Logged from file dns_common_lexicon.py, line 45
Logged from file dns_common_lexicon.py, line 61
Logged from file dns_common_lexicon.py, line 67
Logged from file dns_common.py, line 316
Logged from file dns_common.py, line 64
Logged from file eff.py, line 60
Logged from file eff.py, line 73
Logged from file error_handler.py, line 105
Logged from file error_handler.py, line 110
Logged from file error_handler.py, line 87
Logged from file hooks.py, line 248
Logged from file main.py, line 1071
Logged from file main.py, line 1075
Logged from file main.py, line 1189
Logged from file ops.py, line 122
Logged from file ops.py, line 325
Logged from file ops.py, line 338
Logged from file reporter.py, line 55
Logged from file selection.py, line 110
Logged from file selection.py, line 118
Logged from file selection.py, line 123
Logged from file selection.py, line 176
Logged from file selection.py, line 231
Logged from file selection.py, line 310
Logged from file selection.py, line 66
Logged from file standalone.py, line 101
Logged from file standalone.py, line 88
Logged from file standalone.py, line 97
Logged from file standalone.py, line 98
Logged from file storage.py, line 52
Logged from file storage.py, line 59
Logged from file storage.py, line 75
Logged from file util.py, line 56
Logged from file webroot.py, line 165
Logged from file webroot.py, line 186
Logged from file webroot.py, line 187
Logged from file webroot.py, line 204
Logged from file webroot.py, line 223
Logged from file webroot.py, line 234
Logged from file webroot.py, line 235
Logged from file webroot.py, line 237
Logged from file webroot.py, line 91
```
* Reapply #6667
* Make setuptools delegates tests execution to pytest, like in acme module.
* Clean handlers at each tearDown to avoid memory leaks.
* Update changelog
I observed that the current set of oldest requirements do not correspond to any environment, except the specific Xenial image in Travis CI (and standard Xenial containers will also fail).
It is because the requirements make cryptography and requests fail against standard libraries available in the typical Linux distributions that are targeted by the oldest requirements approach (Centos 6, Centos 7, Xenial, Jessie).
This PR fixes that, by aligning the minimal version requirements of cryptography and requests to the maximal versions that are available on Centos 6. Centos 7, Jessie and Xenial stay unusable with oldest requirements for other reasons, but at least one old and supported Linux distribution is able to run the tests with oldest requirements out of the box.
A test is also corrected to match the expected error message that old versions of urllib3 will raise.
* Remove unneeded sys import.
Once upon a time we needed this in some of these setup.py files because we were
using sys in the file, but we aren't anymore so let's remove the import.
* use setuptools instead of distutils
* Drop support for EOL Python 2.6
* Use more helpful assertIn/NotIn instead of assertTrue/False
* Drop support for EOL Python 3.3
* Remove redundant Python 3.3 code
* Restore code for RHEL 6 and virtualenv for Py2.7
* Revert pipstrap.py to upstream
* Merge py26_packages and non_py26_packages into all_packages
* Revert changes to *-auto in root
* Update by calling letsencrypt-auto-source/build.py
* Revert permissions for pipstrap.py
* Use pipstrap to install a good version of pip
* Use pytest in cb-auto tests
* Remove nose usage in auto_test.py
* remove nose dev dep
* use pytest in test_tests
* Use pytest in tox
* Update dev dependency pinnings
* remove nose multiprocess lines
* Use pytest for coverage
* Use older py and pytest for old python versions
* Add test for Error.__str__
* pin pytest in oldest test
* Fix tests for DNS-DO plugin on py26
* Work around bug for Python 3.3
* Clarify dockerfile comments
* Pin oldest version of packaged python deps
* Install security extras in oldest tests
* Revert "bump requests requirement to >=2.10 (#4248)"
This reverts commit 402ad8b353.
* Use create=True when patching open on module
* Mention python 3 support in setup.py
* Build universal (py2 and py3 compatible) wheels
* Mention Python 3.3+ support in docs
* we work on python 3.6 too
* stop conditionally pinning mock version in acme
* stop conditionally pinning mock version in certbot
* stop conditionally pinning mock version in apache
* stop conditionally pinning mock version in nginx
* stop conditionally pinning mock version in letshelp
* stop conditionally pinning mock version in compatibility-test
The primary motivation is to avoid a branch, giving bugs one fewer place to hide. But, as a bonus, more people get a more bugfixed version of argparse. (To use the example from the argparse docs, people stuck on Python 3.2.3 can get bugfixes that made it into the stdlib only in 3.2.4.)
* pin requests version in py26-oldest
* Determine requests security deps dynamically
Starting with requests 2.12, pyasn1 and ndg-httpsclient are no longer
needed to inject pyopenssl into urllib3. This change allows us to
determine whether or not these dependencies are required at install
time. If an older version of requests is used, these packages are
still installed. If a new version of requests is used, they are not
reducing the number of dependencies we have.
* Bump requests version in certbot-auto
* Use pkg_resources in activate test
Due to pip's lack of dependency resolution, the change to use
requests[extras] causes errors in acme.util_test because pkg_resources
accurately detects the "missing" dependency.
There isn't a real problem here. The problem comes from a brand new
requests and ancient pyopenssl as well as a unit test for
functionality we plan to remove in our next release. I modified
the unit test to fix the problem for now.
* Use six instead of pkg_resources for test
* Require requests<=2.11.1 in py27-oldest test
If we don't do this, we get test failures for the certbot package
which is actually a good thing! pkg_resources is catching the
unlikely but possible problem I describe in #3803 and erroring out
saying it is missing the necessary dependencies to run certbot.
Good job package resources.
* Undo changes to acme.util_test
When pip-installing any of these packages, pip hit our permissive, any-version "setuptools" dependency first and then ignored all subsequent, more constrained ones, like cryptography's "setuptools>=1.0". See https://github.com/pypa/pip/issues/988. It thus, on a box with setuptools 0.9.8, stuck with that version. Then, at runtime, letsencrypt crashed because pkg_resources couldn't satisfy cryptography's setuptools>=1.0 requirement.
This change lets us pip-install our packages and have it work. We'll need to make sure our direct requirements (all of them) satisfy the more constrained requirements of our dependencies. Yes, it is disgusting.
