Follow-up of bpo-40275.

While investigating a crash on AIX (bpo-40068), I noticed that test_threading crashed because the test imports the logging module, and the logging has a bug on AIX on fork. I created an issue to reduce the number of imports made by "import test.support":

https://bugs.python.org/issue40275
I would prefer to better isolate tests: test_threading should only test the threading module, not the logging module.

Thanks to the hard work of Hai Shi, "import test.support" now imports only 37 modules instead of 171! He split the 3200 lines of Lib/test/support/init.py into new helper submodules: bytecode, import, threading, socket, etc. For example, TESTFN now comes from test.support.os_helper.
Sadly, test.regrtest.save_env still imports asyncio and multiprocessing, and so in practice, running any test using "python -m test (...)" still imports around 233 modules :-(

I measured the number of imports done in practice using the following file, Lib/test/test_sys_modules.py:
----

import unittest
from test import support
import sys

class Tests(unittest.TestCase):
    def test_bug(self):
        modules = sorted(sys.modules)
        print("sys.modules:")
        print("")
        import pprint
        pprint.pprint(modules)
        print("")
        print("len(sys.modules):", len(modules))

def test_main():
    support.run_unittest(Tests)

if __name__ == "__main__":
    test_main()

master:

• ./python -m test test_sys_modules: 233 modules (multiprocessing, asyncio, etc.)
• ./python Lib/test/test_sys_modules.py: 95 modules

3.9:

• ./python -m test test_sys_modules: 232
• ./python Lib/test/test_sys_modules.py: 117

3.5:

• ./python -m test test_sys_modules: 167
• ./python Lib/test/test_sys_modules.py: 151

2.7:

• ./python -m test test_sys_modules: 170
• ./python Lib/test/test_sys_modules.py: 122

If I hack test.libregrtest.runtest to not import test.libregrtest.save_env, test_sys_modules imports only 148 instead of 233 modules.

In general, it's nice to have the following 4 checks:

• multiprocessing.process._dangling
• asyncio.events._event_loop_policy
• urllib.requests._url_tempfiles
• urllib.requests._opener

The problem is that because of these checks, **any** unit test file of the 424 Python test files import asyncio, multiprocessing and urllib. As a result, **any** unit test starts with around 233 imported modules. We are far from an "unit" test, since many modules have side effects.

I wrote PR 22089 to remove these checks. "import test.libregrtest" is reduces from 233 to only 149 imports (on Linux), which is way better.

With PR 22089, test_sys_modules.py of msg376374 imports 152 imports rather than 233. It's better than Python 2.7 which imports 170 modules!

On Windows with current master, the baseline for running anything with 1 import (">>>  import sys; len(sys.modules)") is 35 imported modules.  Adding "import unittest" increases this to 80.  What slightly puzzles me is that running 

import unittest
import sys

class Tests(unittest.TestCase):
    def test_bug(self):
        print("len(sys.modules):", len(sys.modules))

if __name__ == "__main__":
    unittest.main()

increases the number to 90. Perhaps unittest has delayed imports.

The current startup number for IDLE is 162, which can result in a cold startup of several seconds. I am thinking of trying to reduce this by delaying imports of modules that are not immediately used and might never be used.

For tests, I gather that side-effect issues are more important than startup time.

You could save/restore this data only when corresponded modules was imported, like it was done in clear_caches() in refleak.py. For example:

    # Same for Process objects
    def get_multiprocessing_process__dangling(self):
        multiprocessing_process = sys.modules.get('multiprocessing.process')
        if not multiprocessing_process:
            return set()
        # Unjoined process objects can survive after process exits
        multiprocessing_process._cleanup()
        # This copies the weakrefs without making any strong reference
        return multiprocessing_process._dangling.copy()
    def restore_multiprocessing_process__dangling(self, saved):
        multiprocessing_process = sys.modules.get('multiprocessing.process')
        if not multiprocessing_process:
            return
        multiprocessing_process._dangling.clear()
        multiprocessing_process._dangling.update(saved)

You could save/restore this data only when corresponded modules was imported, like it was done in clear_caches() in refleak.py.

It was my first idea, but some large modules like multiprocessing and asyncio are only imported by tested when the test file is imported, whereas save_environment() is called (enter) before the import in libregrtest.runtest.

Yes, but they should be imported after running the test. Note that the
default value in the example was changed from None to set().

Serhiy: "You could save/restore this data only when corresponded modules was imported, like it was done in clear_caches() in refleak.py."

That's a very good idea! I implemented it in PR 24934. But I modified runtest() to use *two* saved_test_environment instance. One before the test module is imported, one after. The one before is needed to check if the import itself has side effect, for example if the module body has side effect. The second is to check if running tests has side effect. The second one is more likely to have modules imported. The first one may miss some bugs, but IMO it's an acceptable trade-off.

New changeset 532e063 by Victor Stinner in branch 'master':
bpo-41718: regrtest saved_test_environment avoids imports (GH-24934)
532e063

New changeset 10417dd by Victor Stinner in branch 'master':
bpo-41718: Reduce libregrtest runtest imports (GH-24980)
10417dd

New changeset 0473fb2 by Victor Stinner in branch 'master':
bpo-41718: libregrtest runtest avoids import_helper (GH-24983)
0473fb2

New changeset 30793e8 by Victor Stinner in branch 'master':
bpo-41718: Disable support.testresult XML output by default (GH-24982)
30793e8

New changeset 9feae41 by Victor Stinner in branch 'master':
bpo-41718: libregrtest avoids importing datetime (GH-24985)
9feae41

len(sys.modules) of msg376374 test_sys_modules:

• Python 3.6: 184
• Python 3.7: 183
• Python 3.8: 221
• Python 3.9: 233
• master: 131

The master branch imports 102 less modules than Python 3.9 (233 => 131)! Almost the half.

asyncio, logging, multiprocessing, etc. are no longer always imported by default.

New changeset d72e8d4 by Victor Stinner in branch 'master':
bpo-41718: subprocess imports grp and pwd on demand (GH-24987)
d72e8d4

New changeset bd9154a by Victor Stinner in branch 'master':
bpo-41718: runpy now imports pkgutil in functions (GH-24996)
bd9154a

New changeset cd27af7 by Victor Stinner in branch 'master':
bpo-41718: Update runpy startup time What's New (GH-24998)
cd27af7

Ok, the most important changes have been merged. Thanks everyone who helped on this large project! See also my summary email on python-dev:
https://mail.python.org/archives/list/python-dev@python.org/thread/I3OQTA3F66NQUN7CH2NHC5XZTO24QCIK/