homepage

Copy attached test_leak.py to Lib/test/ and run:
haypo@selma$ ./python -m test -R 3:30 test_leak
Run tests in parallel using 1 child processes
0:00:00 load avg: 0.67 [1/1/1] test_leak failed
beginning 33 repetitions
123456789012345678901234567890123
.................................
test_leak leaked [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] memory blocks, sum=30
1 test failed:
 test_leak
Total duration: 682 ms
Tests result: FAILURE
It seems like Thread2 constructor creates a dictionary on "self.hashvalue = (...)". It's a dictionary sharing keys, but it seems like the dict table memory is not released whne the bug occurs.
Sadly, I'm only able to reproduce the bug on a 32-bit build. Since the bug depends on exact hash values, maybe the bug exists in 64-bit builds, but is less likely?
I'm not 100% sure if the bug is related to dict or set.
The bug comes from bpo-31217.

If you cannot reproduce the bug on 32-bit build, try to set "FIXED_HASH = 0" in test_leak.py, run "./python -m test -R 3:10 test_leak" multiple times until it fails, then copy the last value of the printed list into hash_values in test_leak.py.

I spent a lot of time in gdb, but I'm not sure anymore that the issue is related to dict shared keys.

It's a really weird bug you found. It depends on the exact values passed to -R:
$ ./python -m test -R 3:6 test_haypoleak
Run tests sequentially
0:00:00 load avg: 0.39 [1/1] test_haypoleak
beginning 9 repetitions
123456789
.........
test_haypoleak leaked [1, 1, 1, 1, 1, 1] memory blocks, sum=6
test_haypoleak failed
$ ./python -m test -R 2:6 test_haypoleak
Run tests sequentially
0:00:00 load avg: 0.36 [1/1] test_haypoleak
beginning 8 repetitions
12345678
........
1 test OK.

Actually, it's not the bug that is weird, but the heuristics in -R seem completely weird now. If I print the alloc and rc deltas:
$ ./python -m test -R 3:6 test_haypoleak
Run tests sequentially
0:00:00 load avg: 0.09 [1/1] test_haypoleak
beginning 9 repetitions
123456789
.........
alloc: [0, 0, 0, 1, 1, 1, 1, 1, 1], rc: [0, 0, 0, 0, 0, 0, 0, 0, 0]
test_haypoleak leaked [1, 1, 1, 1, 1, 1] memory blocks, sum=6
test_haypoleak failed
$ ./python -m test -R 2:6 test_haypoleak
Run tests sequentially
0:00:00 load avg: 0.10 [1/1] test_haypoleak
beginning 8 repetitions
12345678
........
alloc: [0, 0, 0, 1, 1, 1, 1, 1], rc: [0, 0, 0, 0, 0, 0, 0, 0]
1 test OK.

s/completely weird/broken/
$ ./python -m test -R 2:20 test_haypoleak
Run tests sequentially
0:00:00 load avg: 0.10 [1/1] test_haypoleak
beginning 22 repetitions
1234567890123456789012
......................
alloc: [0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], rc: [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
1 test OK.

To come back to this bug, it still fails if add a __slots__, so I don't think it's related to shared key dictionaries.

Antoine: "Actually, it's not the bug that is weird, but the heuristics in -R seem completely weird now. If I print the alloc and rc deltas:"
I changed it recently to reduce false alarms: see bpo-30776.
Antoine: "To come back to this bug, it still fails if add a __slots__, so I don't think it's related to shared key dictionaries."
Oh ok, in that case, I will move back to bpo-31217. I close this issue as a duplicate.