Thanks for the detailed report, thinking about what information might be useful, and including it directly in your email.
On 2023-01-28, Roger Marsh <[email protected]> wrote: > Package lmdb-0.9.29 was installed; and py-lmdb, as a --user install, from > PyPi so lmdb can be used from Python3.9. py-lmdb bundles its own copy of LMDB and uses it by default, or you can set extra flags to use the system version. Normally the bundled one is ok, but the LMDB library assumes an OS with coherent file and mmap access (often referred to as 'unified buffer cache') which OpenBSD lacks. The LMDB port (and OpenLDAP) are patched to force MDB_WRITEMAP so all io goes through mmap. This has some drawbacks (and some advantages in certain circumstances) but as far as we know it's the only way we can run it. Obviously the copy bundled in py-lmdb doesn't have this patch. "Use a writeable memory map unless MDB_RDONLY is set. This is faster and uses fewer mallocs, but loses protection from application bugs like wild pointer writes and other bad updates into the database. Incompatible with nested transactions. Do not mix processes with and without MDB_WRITEMAP on the same environment. This can defeat durability (mdb_env_sync etc)." (If I understand correctly, LMDB normally writes through file access, reads through mmap). > Behaviour seemed erratic, including segmentation faults, but for a > while it seemed possible to avoid these by using alternative methods > supported by py-lmdb. Yes this sounds about the expected behaviour for the UBC problem. You can try the port I've just posted to ports@ which uses the system lmdb library. https://marc.info/?l=openbsd-ports&m=167493896320807&w=2 Alternatively set the flags from MAKE_ENV in that port while building py-lmdb manually.
