[FabianIsensee]

Hi Francesc,

a follow-up to #765, from the look at the merged version you asked for. The squash resolved the conflicts with #766 as intended (_store_buffer() decides on the GIL with _can_compress_without_gil(), and errors go through _raise_callback_error()), except in SChunk(data=...): there the resolution kept main's loop, which releases the input buffer only when c-blosc2 returns an error. When _store_buffer() raises, for a chunk larger than BLOSC2_MAX_BUFFERSIZE, the buffer stays exported for good:

import blosc2

data = bytearray(2**31 - 1)
try:
    blosc2.SChunk(chunksize=2**31 - 1, data=data)  # ValueError: Cannot compress more than 2147483615 bytes into one chunk
except ValueError:
    pass
data.append(0)  # main: BufferError: Existing exports of data: object cannot be re-sized (and the 2 GiB stay pinned)

The loop is back in a try/finally, as on the PR branch.

test_data_writes_reach_prefilter ran with the GIL held only, because Python prefilters crashed with set_releasegil(True) before #766. Now that #766 keeps the GIL for them, it runs with the GIL released too, which checks that _store_buffer() follows that rule.

Tests: there is no test for the leak itself, as it needs a 2 GiB buffer; the snippet above raises the BufferError on main and resizes the bytearray with this PR. Full suite: 11001 passed; the 2 b2view http tests fail without aiohttp, as on main.

Merge order: none needed. It merges cleanly with each of the open #769-#775, including #772, which changes _store_buffer() itself.

I used Claude Code (Anthropic) to check the merged version, write the change and run the tests; I reviewed the result.

Best,
Fabian

🤖 Generated with Claude Code