test_threading.ThreadJoinOnShutdown.test_reinit_tls_after_fork() does crash on AIX: bpo-40068. One of the issue is that logging does crash at fork: bpo-40091. The strange thing is that test_threading doesn't use logging. I'm quite sure that logging is imported through test.support.

"import test.support" imports not less than 171... That's quite "heavy". It includes "heavy" modules like concurrent.futures, asyncio or multiprocessing.

It's maybe time to slice again test.support until submodules to reduce the number of imports to the bare minimum. For example, "import test.support" imports asyncio, whereas it's only used by very few tests. Maybe we should add "test.support.asyncioutils".

$ ./python
Python 3.9.0a5+ (heads/pycore_interp:a1ff2c5cf3, Apr 13 2020, 12:27:13) 
>>> import sys
>>> import test

>>> before=set(sys.modules)
>>> import test.support
>>> after=set(sys.modules)

>>> len(after - before)
171

>>> import pprint
>>> pprint.pprint(sorted(after - before))
['__mp_main__',
 '_asyncio',
 '_bisect',
 '_blake2',
 '_bz2',
 '_collections',
 '_compat_pickle',
 '_compression',
 '_contextvars',
 '_datetime',
 '_elementtree',
 '_functools',
 '_hashlib',
 '_heapq',
 '_lzma',
 '_opcode',
 '_operator',
 '_pickle',
 '_posixsubprocess',
 '_queue',
 '_random',
 '_sha3',
 '_sha512',
 '_socket',
 '_sre',
 '_ssl',
 '_string',
 '_struct',
 '_sysconfigdata_d_linux_x86_64-linux-gnu',
 '_weakrefset',
 'argparse',
 'array',
 'asyncio',
 'asyncio.base_events',
 'asyncio.base_futures',
 'asyncio.base_subprocess',
 'asyncio.base_tasks',
 'asyncio.constants',
 'asyncio.coroutines',
 'asyncio.events',
 'asyncio.exceptions',
 'asyncio.format_helpers',
 'asyncio.futures',
 'asyncio.locks',
 'asyncio.log',
 'asyncio.protocols',
 'asyncio.queues',
 'asyncio.runners',
 'asyncio.selector_events',
 'asyncio.sslproto',
 'asyncio.staggered',
 'asyncio.streams',
 'asyncio.subprocess',
 'asyncio.tasks',
 'asyncio.transports',
 'asyncio.trsock',
 'asyncio.unix_events',
 'base64',
 'binascii',
 'bisect',
 'bz2',
 'collections',
 'collections.abc',
 'concurrent',
 'concurrent.futures',
 'concurrent.futures._base',
 'contextlib',
 'contextvars',
 'copy',
 'copyreg',
 'datetime',
 'difflib',
 'dis',
 'email',
 'email.base64mime',
 'email.charset',
 'email.encoders',
 'email.errors',
 'email.header',
 'email.quoprimime',
 'enum',
 'errno',
 'faulthandler',
 'fnmatch',
 'functools',
 'gc',
 'gettext',
 'glob',
 'grp',
 'gzip',
 'hashlib',
 'heapq',
 'importlib',
 'importlib._bootstrap',
 'importlib._bootstrap_external',
 'importlib.abc',
 'importlib.machinery',
 'importlib.util',
 'inspect',
 'itertools',
 'keyword',
 'linecache',
 'locale',
 'logging',
 'logging.handlers',
 'lzma',
 'math',
 'multiprocessing',
 'multiprocessing.context',
 'multiprocessing.process',
 'multiprocessing.reduction',
 'nntplib',
 'opcode',
 'operator',
 'pickle',
 'platform',
 'pprint',
 'pwd',
 'pyexpat',
 'pyexpat.errors',
 'pyexpat.model',
 'queue',
 'quopri',
 'random',
 're',
 'reprlib',
 'resource',
 'select',
 'selectors',
 'shutil',
 'signal',
 'socket',
 'sre_compile',
 'sre_constants',
 'sre_parse',
 'ssl',
 'string',
 'struct',
 'subprocess',
 'sysconfig',
 'tempfile',
 'test.support',
 'test.support.testresult',
 'threading',
 'token',
 'tokenize',
 'traceback',
 'types',
 'typing',
 'typing.io',
 'typing.re',
 'unittest',
 'unittest.async_case',
 'unittest.case',
 'unittest.loader',
 'unittest.main',
 'unittest.result',
 'unittest.runner',
 'unittest.signals',
 'unittest.suite',
 'unittest.util',
 'urllib',
 'urllib.error',
 'urllib.response',
 'warnings',
 'weakref',
 'xml',
 'xml.etree',
 'xml.etree.ElementPath',
 'xml.etree.ElementTree',
 'zlib']

> "import test.support" imports not less than 171... That's quite "heavy".

If we split some "heavy" modules to xxxutils, what benefits  will we make in fact(more exact module importing behavior)?

+1 for splitting test.support on several submodules!

Some imports can also be lazy.

asyncio is now imported in unittest. Removing direct import from test.support does not help.

> asyncio is now imported in unittest. Removing direct import from test.support does not help.

Oh, thanks, you are right. Looks like we need check which submodules should be splitted.

logging is also imported in unittest and indirectly in asyncio.

These three PRs combined reduce the number of imported modules from 179 to 105 (not including 24 modules imported by the bare interpreter) and reduce the hot import time from 160 ms to 80 ms.

> These three PRs combined reduce the number of imported modules from 179 to 105 (not including 24 modules imported by the bare interpreter) and reduce the hot import time from 160 ms to 80 ms.

Good job, serhiy. I tested your patches in my vm, the number of importing module have been reduced. Could you paste your performance benchmark test? I got a no siginificant change :( when I run: 
$ ./python -m pyperf timeit --compare-to python3.9 "from test import support" -p 100
python3.9: ..................................................................................................... 820 ns +- 206 ns

python: ..................................................................................................... 809 ns +- 201 ns

Mean +- std dev: [python3.9] 820 ns +- 206 ns -> [python] 809 ns +- 201 ns: 1.01x faster (-1%)
Not significant!

I used -X importtime.

Your benchmark measures the performance of look up in sys.modules.

> I used -X importtime.
copy that, thanks.

In master branch:
import time:      5011 |     133562 | test.support

After apply serhiy's patches:
import time:      4863 |      66940 | test.support

What is the practical impact on the time taken for test runs? How much does the total time for a test run reduce as a result of refactoring test.support to minimise imports?

Reducing the import time is not a goal, it is just a measurable side effect. The goal is to reduce the number of imported modules unneeded for the particular test. Importing every module can have side effects which affects the tested behavior. It would be nice to minimize it.

Vinay Sajip: "What is the practical impact on the time taken for test runs?"

My first concern is that our "unit tests" are not "unit" anymore. test_threading should be restricted to test the threading module: it should not test "indirectly" the logging module.

If someone wants to test how logging behaves on fork, test_logging should get a new test. (I didn't check if it already has such test.)

New changeset 3c8a5b459d68b4337776ada1e04f5b33f90a2275 by Serhiy Storchaka in branch 'master':
bpo-40275: Avoid importing asyncio in test.support (GH-19600)
https://github.com/python/cpython/commit/3c8a5b459d68b4337776ada1e04f5b33f90a2275

New changeset 16994912c93e8e5db7365d48b75d67d3f70dd7b2 by Serhiy Storchaka in branch 'master':
bpo-40275: Avoid importing socket in test.support (GH-19603)
https://github.com/python/cpython/commit/16994912c93e8e5db7365d48b75d67d3f70dd7b2

New changeset 515fce4fc4bb0d2db97b17df275cf90640017f56 by Serhiy Storchaka in branch 'master':
bpo-40275: Avoid importing logging in test.support (GH-19601)
https://github.com/python/cpython/commit/515fce4fc4bb0d2db97b17df275cf90640017f56

New changeset 66abe98a816de84f89e2de4aa78cf09056227c25 by Hai Shi in branch 'master':
bpo-40275: Move requires_hashdigest() to test.support.hashlib_helper (GH-19716)
https://github.com/python/cpython/commit/66abe98a816de84f89e2de4aa78cf09056227c25

New changeset bfb1cf44658934cbcd9707fb717d6770c78fbeb3 by Serhiy Storchaka in branch 'master':
bpo-40275: Move transient_internet from test.support to socket_helper (GH-19711)
https://github.com/python/cpython/commit/bfb1cf44658934cbcd9707fb717d6770c78fbeb3

New changeset 2935e65c36fab1989bbda19db72c035ea22b044b by Victor Stinner in branch 'master':
bpo-40275: Fix name error in support.socket_helper (GH-19825)
https://github.com/python/cpython/commit/2935e65c36fab1989bbda19db72c035ea22b044b

New changeset 975408c065b645e7d717546b0d744415abb45cd1 by Hai Shi in branch 'master':
bpo-40275: test.support imports lazily locale import (GH-19761)
https://github.com/python/cpython/commit/975408c065b645e7d717546b0d744415abb45cd1

New changeset 7443d42021d433da0497f8ba651daa47e7dc1991 by Hai Shi in branch 'master':
bpo-40275: Import locale module lazily in gettext (GH-19905)
https://github.com/python/cpython/commit/7443d42021d433da0497f8ba651daa47e7dc1991

Hi, folks.
If there have no other imports should be improved, I suggest to close this bpo ;)

While the number of imports reduced a lot, "import test.support" still imports 98 modules. That's a lot! IMO we can reduce it a little bit more :-)

Examples of imports which could be lazy:

* faulthandler: move it into start_threads() and disable_faulthandler()
* shutil: move it into rmtree()
* bz2: move it into @requires_bz2 decorator. Is it possible? Same for zlib, gzip and lzma?


There are also tons of functions, it may be time to create even more submodules:

* support.threading_helper:

  * threading_setup()/threading_cleanup()
  * reap_threads()
  * wait_thread_exit()
  * join_thread()
  * catch_threading_exception context manager
  * start_threads()

* support.process_helper:
  
  * wait_process()
  * reap_children()
  * SuppressCrashReport
  * PIPE_MAX_SIZE (is it the best place for that one?)
  * args_from_interpreter_flags()
  * optim_args_from_interpreter_flags()
  * PythonSymlink

* support.file_helper: 

  * rmtree() and the related private functions
  * unlink()
  * rmdir()
  * FS_NONASCII
  * TESTFN_UNICODE, TESTFN_NONASCII, TESTFN_UNENCODABLE, TESTFN_UNDECODABLE
  * SAVEDCWD (not sure about this one)
  * TESTFN_ENCODING <= this one can be removed, it's just sys.getfilesystemencoding()
  * temp_dir()
  * change_cwd()
  * temp_cwd()
  * temp_umask()
  * create_empty_file()
  * make_bad_fd()
  * EnvironmentVarGuard (not sure about this one)
  * can_symlink(), skip_unless_symlink()
  * can_xattr(), skip_unless_xattr()
  * fs_is_case_insensitive()
  * fd_count()
  * FakePath

* support.import_helper:

  * import_module(), _ignore_deprecated_imports()
  * import_fresh_module(), _save_and_remove_module(), _save_and_block_module()
  * unload()
  * make_legacy_pyc()
  * forget()
  * CleanImport
  * DirsOnSysPath
  * modules_setup(), modules_cleanup()

* support.warnings_helper:

  * check_syntax_warning()
  * ignore_warnings()
  * WarningsRecorder
  * check_warnings(), _filterwarnings()
  * check_no_warnings()
  * check_no_resource_warning()

Move also move the following functions to socket_helper:

  * SOCK_MAX_SIZE
  * open_urlresource()
  * TransientResource
  * time_out
  * socket_peer_reset
  * ioerror_peer_reset

Serhiy: What do you think of these proposed submodules?

OK, I continue to reduce the import module in test.supports.

After PR20128, the import modules from 132 to 123.

New changeset 372fa3ead584876a975a61936b376259be636d27 by Hai Shi in branch 'master':
bpo-40275: lazy import modules in test.support (GH-20128)
https://github.com/python/cpython/commit/372fa3ead584876a975a61936b376259be636d27

I think that there is no much benefit in avoiding to import modules which are imported in libregrtest. In particular threading, subprocess, tempdir, os, fnmatch, etc. You should minimize import of test.libregrtest + test.support, not just test.support. BTW, libregrtests imports now much more modules than test.support.

Also, some modules, like bz2 are too small and do not have much dependencies. You will not save much on importing them lazily. On other hand, lazy import have its cost, so the real benefit will be even smaller.

Serhiy:
> I think that there is no much benefit in avoiding to import modules which are imported in libregrtest.

Well, we should also enhance libregrtest in this case :-)

> You should minimize import of test.libregrtest + test.support, not just test.support.

Good idea.

> Also, some modules, like bz2 are too small and do not have much dependencies. You will not save much on importing them lazily. On other hand, lazy import have its cost, so the real benefit will be even smaller.

FYI bz2 and lzma can me some headaches when experimenting isolated subinterpreters while I was testing modules which don't use them. bz2 and lzma are not compatible with subintepreters and so caused crashes.

IMO in an ideal world, running test_xxx should start with an empty sys.modules and only imports what it needs. The problem are also side effects, not only memory footprint.

I created this issue because test_threading crashed on AIX because of a bug at fork in the logging module. Except that test_threading and threading modules *don't* importing logging. It's non obvious to get a crash in logging while testing the threading module. For me, these are tests are no longer "unit" tests, but more "integration" tests if we test indirectly "half of the stdlib" (I exaggerate on purpose ;-)).

New changeset a3ec3ad9e20e7d9ed148d4cfbd22aebec608b42a by Hai Shi in branch 'master':
bpo-40275: More lazy imports in test.support (GH-20131)
https://github.com/python/cpython/commit/a3ec3ad9e20e7d9ed148d4cfbd22aebec608b42a

New changeset e80697d687b610bd7fb9104af905dec8f0bc55a7 by Hai Shi in branch 'master':
bpo-40275: Adding threading_helper submodule in test.support (GH-20263)
https://github.com/python/cpython/commit/e80697d687b610bd7fb9104af905dec8f0bc55a7

New changeset b0461e19b5ecb2d89917b23efb5ce1048fab1b22 by Victor Stinner in branch 'master':
bpo-40275: test.support.check_impl_detail() uses sys.implementation (GH-20468)
https://github.com/python/cpython/commit/b0461e19b5ecb2d89917b23efb5ce1048fab1b22

Attached count_imports.py script counts how many modules are imported by "import test.support".

With PR 20471, we are down to 74 modules. It's way better than 171 (when I created this issue)!

New changeset aa890630bc877c73ec806c9982d8b7b2e7019723 by Victor Stinner in branch 'master':
bpo-40275: test.support imports subprocess lazily (GH-20471)
https://github.com/python/cpython/commit/aa890630bc877c73ec806c9982d8b7b2e7019723

See also bpo-40799: Create Lib/_pydecimal.py file to optimize "import datetime" when _decimal is available.

New changeset 24bddc1b3b58f6899b2d412e51b37f68536e4fe2 by Hai Shi in branch 'master':
 bpo-40275: Remove test.support.TESTFN_ENCODING (GH-20482)
https://github.com/python/cpython/commit/24bddc1b3b58f6899b2d412e51b37f68536e4fe2

New changeset 753643205a28531fd43ef36b40b86282ae6956a7 by Victor Stinner in branch 'master':
bpo-40275: Fix test.support.threading_helper (GH-20488)
https://github.com/python/cpython/commit/753643205a28531fd43ef36b40b86282ae6956a7

New changeset 0d00b2a5d74390da7bbeff7dfa73abf2eb46124a by Hai Shi in branch 'master':
bpo-40275: Add os_helper submodule in test.support (GH-20765)
https://github.com/python/cpython/commit/0d00b2a5d74390da7bbeff7dfa73abf2eb46124a

New changeset 7f888c7ef905842bf7739cc03bd20398329951b5 by Hai Shi in branch 'master':
bpo-40275: Add import_helper submodule in test.support (GH-20794)
https://github.com/python/cpython/commit/7f888c7ef905842bf7739cc03bd20398329951b5

New changeset bdfe9b633a171522bc6cdb7c8aa1215cbd119b59 by Victor Stinner in branch 'master':
bpo-40275: test.supports imports lazily fnmatch, glob, struct (GH-20810)
https://github.com/python/cpython/commit/bdfe9b633a171522bc6cdb7c8aa1215cbd119b59

New changeset 10e6506aa8261aacc89b49e629ae1c927fa5151c by Hai Shi in branch 'master':
bpo-40275: Add warnings_helper submodule in test.support (GH-20797)
https://github.com/python/cpython/commit/10e6506aa8261aacc89b49e629ae1c927fa5151c

I'm no longer sure that it's worth it to create support.process_helper. I only see these functions which would be untroversial for process_helper:

* wait_process()
* reap_children()
* args_from_interpreter_flags()
* optim_args_from_interpreter_flags()

PythonSymlink might be moved to support.script_helper instead.

For now, I suggest to leave all these functions in support/__init__.py.

I propose to start updating tests to use directly support.import_helper and support.os_helper. Hai: do you want to work on that? Maybe start by focusing on import_helper (but if you modify a test, check for also symbols imported from os_helper).

When you modify tests, I suggest to remove "from .import_helper import ..." and "from .os_helper import ..." in support/__init__.py and then run tests, to ensure that you updated all imports of the modified tests.

Try to not create giant PRs modifying 100+ files. For example, limit a PR to 20 test files.

When PR 20812 will be merged and import_helper and os_helper will no longer be imported  by support, "import test.support" done after "import unittest"  will only import 17 modules which is way better than the situation in April!

I propose to start updating tests to use directly support.import_helper and support.os_helper. Hai: do you want to work on that? Maybe start by focusing on import_helper (but if you modify a test, check for also symbols imported from os_helper).

> of course, I will do it.

Try to not create giant PRs modifying 100+ files. For example, limit a PR to 20 test files.

> forgive me again. I will create some PRs of appropriate size.

New changeset 311110abcd8ab648dbf1803e36a8ba5d93fa019b by Victor Stinner in branch 'master':
bpo-40275: Move TransientResource to test_urllib2net (GH-20812)
https://github.com/python/cpython/commit/311110abcd8ab648dbf1803e36a8ba5d93fa019b

See also bpo-41006: "Reduce number of modules imported by runpy".

New changeset f7ba40b505989495c3585ed782070bdae56330ad by Hai Shi in branch 'master':
bpo-40275: Use new test.support helper submodules in tests (GH-20849)
https://github.com/python/cpython/commit/f7ba40b505989495c3585ed782070bdae56330ad

New changeset 06a40d735939fd7d5cb77a68a6e18299b6484fa5 by Hai Shi in branch 'master':
bpo-40275: Use new test.support helper submodules in tests (GH-20824)
https://github.com/python/cpython/commit/06a40d735939fd7d5cb77a68a6e18299b6484fa5

New changeset 847f94f47b104aec678d1d2a2d8fe23d817f375e by Hai Shi in branch 'master':
bpo-40275: Use new test.support helper submodules in tests (GH-21151)
https://github.com/python/cpython/commit/847f94f47b104aec678d1d2a2d8fe23d817f375e

New changeset 3ddc634cd5469550c0c2dc5a6051a70739995699 by Hai Shi in branch 'master':
bpo-40275: Use new test.support helper submodules in tests (GH-21219)
https://github.com/python/cpython/commit/3ddc634cd5469550c0c2dc5a6051a70739995699

New changeset 0c4f0f3b29d84063700217dcf90ad6860ed71c70 by Hai Shi in branch 'master':
bpo-40275: Use new test.support helper submodules in tests (GH-21169)
https://github.com/python/cpython/commit/0c4f0f3b29d84063700217dcf90ad6860ed71c70

New changeset 883bc638335a57a6e6a6344c2fc132c4f9a0ec42 by Hai Shi in branch 'master':
bpo-40275: Use new test.support helper submodules in tests (GH-21314)
https://github.com/python/cpython/commit/883bc638335a57a6e6a6344c2fc132c4f9a0ec42

New changeset a089d21df1ea502b995d8e8a3bcc937cce030802 by Hai Shi in branch 'master':
bpo-40275: Use new test.support helper submodules in tests (GH-21315)
https://github.com/python/cpython/commit/a089d21df1ea502b995d8e8a3bcc937cce030802

New changeset deb016224cc506503fb05e821a60158c83918ed4 by Hai Shi in branch 'master':
bpo-40275: Use new test.support helper submodules in tests (GH-21317)
https://github.com/python/cpython/commit/deb016224cc506503fb05e821a60158c83918ed4

New changeset 96a6a6d42be272a27562d98549bbffc0d1854669 by Hai Shi in branch 'master':
bpo-40275: Use new test.support helper submodules in tests (GH-21412)
https://github.com/python/cpython/commit/96a6a6d42be272a27562d98549bbffc0d1854669

New changeset a7f5d93bb6906d0f999248b47295d3a59b130f4d by Hai Shi in branch 'master':
bpo-40275: Use new test.support helper submodules in tests (GH-21449)
https://github.com/python/cpython/commit/a7f5d93bb6906d0f999248b47295d3a59b130f4d

New changeset bb0424b122e3d222a558bd4177ce37befd3e0347 by Hai Shi in branch 'master':
bpo-40275: Use new test.support helper submodules in tests (GH-21451)
https://github.com/python/cpython/commit/bb0424b122e3d222a558bd4177ce37befd3e0347

New changeset 4660597b51b3d14ce6269d0ed865ab7e22c6ae1f by Hai Shi in branch 'master':
bpo-40275: Use new test.support helper submodules in tests (GH-21448)
https://github.com/python/cpython/commit/4660597b51b3d14ce6269d0ed865ab7e22c6ae1f

New changeset 604bba1f8fe32479c89b9824b2231cc4480dd110 by Hai Shi in branch 'master':
bpo-40275: Use new test.support helper submodules in tests (GH-21452)
https://github.com/python/cpython/commit/604bba1f8fe32479c89b9824b2231cc4480dd110

New changeset c7decc27d529c04a4e6b2922e3f3f9419b920f63 by Hai Shi in branch 'master':
bpo-40275: Use new test.support helper submodules in tests (GH-21727)
https://github.com/python/cpython/commit/c7decc27d529c04a4e6b2922e3f3f9419b920f63

New changeset 79bb2c93f2d81702fdf1f93720369e18a76b7d1a by Hai Shi in branch 'master':
bpo-40275: Use new test.support helper submodules in tests (GH-21743)
https://github.com/python/cpython/commit/79bb2c93f2d81702fdf1f93720369e18a76b7d1a

New changeset 598a951844122678de2596dbc1e0e09e2be65fd2 by Hai Shi in branch 'master':
bpo-40275: Use new test.support helper submodules in tests (GH-21764)
https://github.com/python/cpython/commit/598a951844122678de2596dbc1e0e09e2be65fd2

New changeset fcce8c649a14f7a81fae82f9f203bb5b5ee0c205 by Hai Shi in branch 'master':
bpo-40275: Use new test.support helper submodules in tests (GH-21772)
https://github.com/python/cpython/commit/fcce8c649a14f7a81fae82f9f203bb5b5ee0c205

New changeset d94af3f7ed98e6bfb4bf5f918f464b5e23d3ed1b by Hai Shi in branch 'master':
bpo-40275: Remove test helpers aliases in test.support (GH-21771)
https://github.com/python/cpython/commit/d94af3f7ed98e6bfb4bf5f918f464b5e23d3ed1b

New changeset c6f282f3b1cb6da6febc3b8b6d2dc1ef714dbbf7 by Hai Shi in branch 'master':
bpo-40275: Use new test.support helper submodules in tests (GH-21785)
https://github.com/python/cpython/commit/c6f282f3b1cb6da6febc3b8b6d2dc1ef714dbbf7

PR-21771 "Remove test helpers aliases in test.support" was just merged.  It needs an immediate followup to document the new locations of constants and functions.  The now dead entries in test.suppport doc must be moved into new support module docs, with whatever revisions. As of this moment, the module index has
test
  test.support
  test.support.script_helper
So temp_dir, for instance, is still documented as part of test.support, which it no longer is, instead of the undocumented test.support.os_helper, where it now exclusively resides.

Hi, terry. When creating the new test helpers, the docs have been updated.
such as: https://docs.python.org/3.10/library/test.html#module-test.support.os_helper

PR21771 has broken a considerable amount of buildbopts so we would need to revert it if is not fixed in 24 hours

>> This PR has broken a considerable amount of build bots, so we would need to revert it if is not fixed in 24 hours

Whops, sorry I just saw that this is being fixed here #21785, apologies then!

Sorry, I forgot that for the website, 3.x means 3.8, which has only 1 x_helper module.  3.9 has 3 and 3.10 has 7.

The buildbots were fixed with PR-21785.  '#' prefixes an issue on bpo, a PR on github.  Confusing, especially now that PR #s are catching up to issue #s.

# Confusing, especially now that PR #s are catching up to issue #s.
Maybe we need redefine the matching rule(I am not sure about it).

test__osx_support and test_selectors are broken, examples:

https://github.com/python/cpython/pull/21806/checks?check_run_id=966438645

2020-08-10T12:04:20.7613070Z ======================================================================
2020-08-10T12:04:20.7613710Z ERROR: test__remove_unsupported_archs (test.test__osx_support.Test_OSXSupport)
2020-08-10T12:04:20.7616560Z ----------------------------------------------------------------------
2020-08-10T12:04:20.7617240Z Traceback (most recent call last):
2020-08-10T12:04:20.7617860Z   File "/Users/runner/work/cpython/cpython/Lib/test/test__osx_support.py", line 23, in setUp
2020-08-10T12:04:20.7618560Z     self.env = test.support.EnvironmentVarGuard()
2020-08-10T12:04:20.7621650Z AttributeError: module 'test.support' has no attribute 'EnvironmentVarGuard'

and

2020-08-10T12:01:40.0736200Z ======================================================================
2020-08-10T12:01:40.0738440Z ERROR: test_register_bad_fd (test.test_selectors.KqueueSelectorTestCase)
2020-08-10T12:01:40.0742700Z ----------------------------------------------------------------------
2020-08-10T12:01:40.0744640Z Traceback (most recent call last):
2020-08-10T12:01:40.0747460Z   File "/Users/runner/work/cpython/cpython/Lib/test/test_selectors.py", line 539, in test_register_bad_fd
2020-08-10T12:01:40.0753630Z     bad_f = support.make_bad_fd()
2020-08-10T12:01:40.0757650Z AttributeError: module 'test.support' has no attribute 'make_bad_fd'

Hai Shi: Would you mind to investigate this issue?

> Hai Shi: Would you mind to investigate this issue?
Oh, sure, I will check the test cases again.

New changeset 490c5426b1b23831d83d0c6b269858fb98450889 by Hai Shi in branch 'master':
bpo-40275: Fix failed test cases by using test helpers (GH-21811)
https://github.com/python/cpython/commit/490c5426b1b23831d83d0c6b269858fb98450889

Update.

"import test.support" now imports 37 modules, instead of 171 previously. It's way better!

23:26:20 vstinner@apu$ ./python
Python 3.10.0a0 (heads/master:598a951844, Aug  7 2020, 17:24:10) 
[GCC 10.2.1 20200723 (Red Hat 10.2.1-1)] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import sys
>>> import test
>>> before=set(sys.modules)
>>> import test.support
>>> after=set(sys.modules)
>>> len(after - before)
37
>>> import pprint; pprint.pprint(sorted(after - before))
['_datetime',
 '_elementtree',
 '_heapq',
 '_sysconfigdata_d_linux_x86_64-linux-gnu',
 '_testcapi',
 '_weakrefset',
 'argparse',
 'copy',
 'datetime',
 'difflib',
 'fnmatch',
 'gettext',
 'heapq',
 'math',
 'pprint',
 'pyexpat',
 'pyexpat.errors',
 'pyexpat.model',
 'signal',
 'sysconfig',
 'test.support',
 'test.support.testresult',
 'traceback',
 'unittest',
 'unittest.case',
 'unittest.loader',
 'unittest.main',
 'unittest.result',
 'unittest.runner',
 'unittest.signals',
 'unittest.suite',
 'unittest.util',
 'weakref',
 'xml',
 'xml.etree',
 'xml.etree.ElementPath',
 'xml.etree.ElementTree']


Shorter list if imports made by unittest are ignored:

23:27:42 vstinner@apu$ ./python
Python 3.10.0a0 (heads/master:598a951844, Aug  7 2020, 17:24:10) 
[GCC 10.2.1 20200723 (Red Hat 10.2.1-1)] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import sys, unittest, test
>>> before=set(sys.modules)
>>> before=set(sys.modules); import test.support; after=set(sys.modules)
>>> len(after) - len(before)
17
>>> import pprint; pprint.pprint(sorted(after - before))
['_datetime',
 '_elementtree',
 '_sysconfigdata_d_linux_x86_64-linux-gnu',
 '_testcapi',
 'copy',
 'datetime',
 'math',
 'pyexpat',
 'pyexpat.errors',
 'pyexpat.model',
 'sysconfig',
 'test.support',
 'test.support.testresult',
 'xml',
 'xml.etree',
 'xml.etree.ElementPath',
 'xml.etree.ElementTree']

Hi, victor:
  About https://bugs.python.org/issue40275#msg369214, shall we continue to improve the libregretest?
1. Running test cases sequential: you have use a mechanism to unload the modules which load in the running test cases stage: https://github.com/python/cpython/blob/master/Lib/test/libregrtest/main.py#L437
2. Running test cases concurrently: It have use subprocess to isolate the module resources, no?
I am not sure about it.

> About https://bugs.python.org/issue40275#msg369214, shall we continue to improve the libregretest?

If someone wants to continue the effort in libregrtest, I suggest to open a new issue. This one already has a long list of commits and messages about test.support.

See test.libregrtest.runtest_mp module for the "multiprocessing" implementation which uses subprocess. It's used when you run tests using -jN option. Almost all buildbots use -jN: see make buildbottest.

If someone wants to continue the effort in libregrtest, I suggest to open a new issue.
+1. And if there have any other test cases blocked by test.support, I suggest this bpo.

oh, sorry, missing a word: I suggest this bpo.->I suggest close this bpo.

Sadly, the work done on this issue is useless until test.regrtest imports less modules as well.

So I created bpo-41718 follow-up issue: "test.regrtest has way too many imports".

I consider that the work is done on test.support, so I close this issue.