support._match_test() uses a nested loop calling fnmatch.fnmatchcase(). This function creates a temporary regular expression object. The cache of the re module works around the performance... if the length of support.match_tests fits into the cache. But when the list is longer, the function becomes dead slow...

def _match_test(test):
    global match_tests
                                                                                                   
    if match_tests is None:
        return True 
    test_id = test.id()                                                                            
                
    for match_test in match_tests:
        if fnmatch.fnmatchcase(test_id, match_test):                                               
            return True
                                                                                                   
        for name in test_id.split("."):                                                            
            if fnmatch.fnmatchcase(name, match_test):                                              
                return True
    return False

Maybe we should build a giant regex matching test_id at each, but cache the regex since support.match_tests can be modified anytime. I implemented this once, but I lost the code :-)

Currently, it's possible to match 3 things:

* test method name: test_exit
* test class name: SysModuleTest
* full test id: test.test_sys.SysModuleTest.test_exit

It's also possible to use "*" joker character in a test name. I would like to keep these convenient CLI.

Example of the performance issue:

$ ./python -m test.bisect --fail-env-changed -o bisect test_asyncio

[+] Iteration 1: run 756 tests/1512
(...)

The first iteration takes forever because it runs tests with support.match_tests which contains 756 patterns.

Workaround to run test.bisect on test_asyncio:

diff --git a/Lib/test/support/__init__.py b/Lib/test/support/__init__.py
index 522804be60..201d0665b2 100644
--- a/Lib/test/support/__init__.py
+++ b/Lib/test/support/__init__.py
@@ -1911,15 +1911,7 @@ def _match_test(test):
     if match_tests is None:
         return True
     test_id = test.id()
-
-    for match_test in match_tests:
-        if fnmatch.fnmatchcase(test_id, match_test):
-            return True
-
-        for name in test_id.split("."):
-            if fnmatch.fnmatchcase(name, match_test):
-                return True
-    return False
+    return (test_id in match_tests)
 
 
 def run_unittest(*classes):


With this patch, running 756 tests takes 12 seconds, as expected.

Oh, by the way, -m test -m TestCaseName doesn't work fully on test_asyncio. It only works if I pass the full test identifier.

With TestCaseName, no test is run:

haypo@selma$ ./python -m test -u all test_asyncio --fail-env-changed -m PollEventLoopTests 
Run tests sequentially
0:00:00 load avg: 6.97 [1/1] test_asyncio
1 test OK.

Total duration: 458 ms
Tests result: SUCCESS


With a full test identifier, a test is run:


haypo@selma$ ./python -m test -u all test_asyncio --fail-env-changed -m test.test_asyncio.test_events.PollEventLoopTests.test_create_datagram_endpoint_sock
Run tests sequentially
0:00:00 load avg: 7.06 [1/1] test_asyncio
Warning -- threading_cleanup() detected 1 leaked threads (count: 1, dangling: 2)
test_asyncio failed (env changed)

1 test altered the execution environment:
    test_asyncio

Total duration: 561 ms
Tests result: ENV CHANGED

Oops. Ignore my previous comment, I was using my temporary workaround! (msg301126)

_match_test1 = re.compile('|'.join(map(fnmatch.translate, match_tests)), re.I).match

def _match_test(test):
    test_id = test.id()
    return bool(_match_test1(test_id) or any(map(_match_test1, test_id.split(".")))

I tested manually Serhiy's PR 4420: it works as expected!

haypo@selma$ ./python -m test test_os --list-cases -m test.test_os.FileTests.test_access
test.test_os.FileTests.test_access

haypo@selma$ ./python -m test test_os --list-cases -m test_access
test.test_os.FileTests.test_access

haypo@selma$ ./python -m test test_os --list-cases -m FileTests
test.test_os.FileTests.test_access
test.test_os.FileTests.test_closerange
test.test_os.FileTests.test_fdopen
test.test_os.FileTests.test_large_read
test.test_os.FileTests.test_open_keywords
test.test_os.FileTests.test_read
test.test_os.FileTests.test_rename
test.test_os.FileTests.test_replace
test.test_os.FileTests.test_symlink_keywords
test.test_os.FileTests.test_write
test.test_os.FileTests.test_write_windows_console

haypo@selma$ ./python -m test test_os --list-cases -m 'FileTest*'
test.test_os.FileTests.test_access
test.test_os.FileTests.test_closerange
test.test_os.FileTests.test_fdopen
test.test_os.FileTests.test_large_read
test.test_os.FileTests.test_open_keywords
test.test_os.FileTests.test_read
test.test_os.FileTests.test_rename
test.test_os.FileTests.test_replace
test.test_os.FileTests.test_symlink_keywords
test.test_os.FileTests.test_write
test.test_os.FileTests.test_write_windows_console

haypo@selma$ ./python -m test test_os --list-cases -m 'test_acc*'
test.test_os.FileTests.test_access
test.test_os.StatAttributeTests.test_access_denied

haypo@selma$ ./python -m test test_os --list-cases -m '*FileTests.test_access'
test.test_os.FileTests.test_access

$ time ./python -m test.bisect --fail-env-changed -o bisect test_asyncio -N1
...
Bisection failed after 2 iterations and 0:00:11

real    0m10,441s
user    0m8,726s
sys     0m0,292s

There is good opportunity for optimizing the regular expression. All cases have the literal 'test.test_asyncio.' prefix, and even longer literal prefixes are common for tens cases.

> There is good opportunity for optimizing the regular expression. All cases have the literal 'test.test_asyncio.' prefix, and even longer literal prefixes are common for tens cases.

I implemented such code long time ago :-)

https://github.com/haypo/hachoir3/blob/master/hachoir/regex/regex.py

I'm not sure that this code works on Python 3, it was written for Python 2 but not only ported to Python 3 partially.

It only supports a subset of regular expressions.

I know :-)

Unfortunately it is not obviously that these optimization actually optimize regular expressions. Due to the implementation particularities sometimes "unoptimized" code is faster than "optimized". I'm going to implement these optimizations in the re module, but it needs to be very careful, and may need changing the matching engine.

I created a list of all test cases using:

./python -m test --list-cases > all_cases

The list contains 29,569 test cases. Sadly, the set().__contains__ matcher of my PR 4421 cannot be taken because test_json produces two test cases called "json". IMHO it's a bug in test_json, all test identifiers should be "fully qualified". test_builtin has a similar issue: it produces identifiers like "builtins.bin" instead of "test.test_builtin.builtins.bin".

If I removed the two "json" from all_cases, "./python -u -m test.bisect -i all_cases -v" takes less than one second before running the first test method, and so it's fast enough.

With PR 4420, it takes around 17 seconds before running the first test. The compilation of the giant regex and matching test identifiers are slow. My PR has the same bad performance when the regex path is taken.

We could split patterns on two parts and create both matchers. Then the final matching function could look like:

return _match_test_patterns is None or test_id in id_set or regex_match(test_id) or any(map(regex_match, test_id.split(".")))


I don't know whether it is worth to do.

New changeset 803ddd8ce22f0de3ab42fb98a225a704c000ef06 by Victor Stinner in branch 'master':
bpo-31324: Optimize support._match_test() (#4421)
https://github.com/python/cpython/commit/803ddd8ce22f0de3ab42fb98a225a704c000ef06

I merged my PR 4421 which is based on Serhiy's PR 4420. Thank you Serhiy for your reviews!

--

Serhiy Storchaka: "We could split patterns on two parts and create both matchers (...) I don't know whether it is worth to do."

My use bisect is to find a regression using the test.bisect tool. According to my tests, the fastest set().__contains__ matcher is always used, except of test_json. If you bisect a bug in test_json, it will work, but just be a little bit slower.

Honestly, I don't think that we can go further in term of optimization. In my tests, it seems like creating the matcher function and filter all tests take less than 1 second, so it's enough for my use case ;-)

Oops, the commit 803ddd8ce22f0de3ab42fb98a225a704c000ef06 broke the GCC job on Travis CI:
---
Run tests sequentially

0:00:00 load avg: 48.91 [  1/404] test_grammar

0:00:00 load avg: 48.91 [  2/404] test_opcodes

0:00:00 load avg: 48.91 [  3/404] test_dict

0:00:00 load avg: 48.91 [  4/404] test_builtin

0:00:00 load avg: 48.91 [  5/404] test_exceptions

0:00:01 load avg: 48.91 [  6/404] test_types

0:00:01 load avg: 48.91 [  7/404] test_unittest

0:00:03 load avg: 48.91 [  8/404] test_doctest

0:00:04 load avg: 48.11 [  9/404] test_doctest2

0:00:04 load avg: 48.11 [ 10/404] test_support

0:00:07 load avg: 48.11 [ 11/404] test___all__

Traceback (most recent call last):

  File "/home/travis/build/python/cpython/Lib/test/__main__.py", line 2, in <module>

    main()

  File "/home/travis/build/python/cpython/Lib/test/libregrtest/main.py", line 585, in main

    Regrtest().main(tests=tests, **kwargs)

  File "/home/travis/build/python/cpython/Lib/test/libregrtest/main.py", line 510, in main

    self._main(tests, kwargs)

  File "/home/travis/build/python/cpython/Lib/test/libregrtest/main.py", line 536, in _main

    self.run_tests()

  File "/home/travis/build/python/cpython/Lib/test/libregrtest/main.py", line 458, in run_tests

    self.run_tests_sequential()

  File "/home/travis/build/python/cpython/Lib/test/libregrtest/main.py", line 379, in run_tests_sequential

    result = runtest(self.ns, test)

  File "/home/travis/build/python/cpython/Lib/test/libregrtest/runtest.py", line 105, in runtest

    support.set_match_tests(ns.match_tests)

  File "/home/travis/build/python/cpython/Lib/test/support/__init__.py", line 1958, in set_match_tests

    _match_test_patterns = tuple(patterns)

TypeError: 'NoneType' object is not iterable
---

Attached PR 4505 fixes it.

New changeset bb11c3c967afaf263e00844d4ab461b7fafd6d36 by Victor Stinner in branch 'master':
bpo-31324: Fix test.support.set_match_tests(None) (#4505)
https://github.com/python/cpython/commit/bb11c3c967afaf263e00844d4ab461b7fafd6d36

New changeset 70b2f8797146a56a6880743424f0bedf4fc30c62 by Victor Stinner in branch '3.6':
[3.6] bpo-31324: Optimize support._match_test() (#4523)
https://github.com/python/cpython/commit/70b2f8797146a56a6880743424f0bedf4fc30c62

New changeset 35d99830f1878867e3964577741d9a2d5a7fc8f7 by Victor Stinner in branch '2.7':
bpo-31324: Optimize support._match_test() (#4523) (#4524)
https://github.com/python/cpython/commit/35d99830f1878867e3964577741d9a2d5a7fc8f7

I tested set_match_tests() on python2.7 with:

python2.7 -m test --list-cases > cases
python2.7 -m test.bisect -i cases

Good news: the set().__contains__ matcher can be used!