Fix Github Actions

[andarut]

The problem

Github Actions fails randomly. On my fork i rerun all jobs for same code many times. And got this:

I think it's very important to fix actions, because without all checks passed branch can't be merged and of course there is no trust to PR. For example my previous PR Fix build C libraries inside kphp. @Danil42Russia rerun checks on my PR 3 times and no one succed. So branch with useful fix can't be merged, because of random. And this is not about code in PR, this is about randomly failed tests. You can check my fork for more (i tested another my PR Add mbstring functions to kphp): my fork.

The path and fix

Linux build

So, there is 5 randomly failing python tests (timeout error):

• test_define_from_config
• test_script_errors
• test_headers_limit
• test_post_limit
• test_query_limit

The cause is that the latest version of requests does not support urllib3 2.0.
This is triggered by versions of requests-toolbelt and urllib3 that were both released in the past few weeks (May 1 and May 4, 2023). This is already an issue in requests-toolbelt. This is refer to urllib3 2.0 bug: urllib3 keeps a reference to that exception to include it in the MaxRetryError exception it will eventually raise. This means the garbage collector will not close the socket on its own, which means the connection will be kept open. So when urllib3 retries, it won't be able to open a connection, because the server is waiting for the previous try to finish. That's a deadlock!

Note! I've tried to rewrite send_http_request function from tests/python/lib/http_server.py with urllib requests. This is wrong way, urllib have different work style. Downgrade urllib much more safety.

And 1 test, randomly failing with RuntimeError: Got bad stat line error

• test_store_fetch_delete

To fix this, i changed tests/python/lib/stats_receiver.py that way, so now when we receive not complete stats key, we just skip it. That's right because after this test fails, even logs shows that stats file is full and right. So sometimes engine don't be on time to write full string, when test runs. But old code already have timeout function for that (wait_next_stats). So i think when we wait for updating stats and got not full key, we must't throw runtime error. Also when get ValueError when unpacking key, this is also reason to wait for engine to write.

And 1 test, randomly failing because of timeout:

• test_job_stack_overflow_error

Test deleted, because of incorrect work with ASAN. ASAN randomly shutdown worker when getting stackoverflow error, so we have timeout.

MacOS build

So, there is 3 failed cpp tests:

• counter_test
• parallel_limit_counter
• parallel_counter

All tests deal with threads. I think that slow syscalls might cause other threads to spawn to take over. Github Actions default mac os is 3 cores machine, so maybe the efficiency cores are too slow to take over and kphp thinks the syscall takes too long. So i disable these tests similar to maximum-test.cpp (which also deals with threads).

Result

Result

[Danil42Russia]

I give my approve, ONLY for running tests.

This does not mean that this PR is ready to merge in master.

Tests passed

[DrDet]

Maybe we'll remove this completely?

[andarut]

Ok, done.