[davidackerman]

0.8.0: faster writes from the contexts the package makes, and the create open timed out. No placement change and no unit table change. It requires tensorstore 0.1.68 or later (was 0.1.60), the first release with file_io_locking.

All of this was measured on 2026-10-23 on real CellMap data, against zarr-python 2.18 and 3.1, on local NVMe and NFS.

fsync and locking. shared_context and capped_context now set file_io_sync: false and file_io_locking: {"mode": "none"}. fsync=True and locking="os" restore tensorstore's defaults.

• fsync cost 6.5 ms per chunk on local disk and nothing on NFS, which commits a file on close. zarr-python and funlib never synced.
• "os" locking cost about 10 ms per chunk on NFS: a 64³ label chunk took 21.7 ms with it and 12.0 ms without, against zarr-python 2's 12.7 ms.
• What "none" keeps and gives up. A whole-unit write stays atomic, as a temporary file renamed into place. Only concurrent read-modify-writes of one unit lose their guarantee, and Behaviour 13 never gave that.
• Why not "lockfile". A lock file left behind by a killed worker blocks every later write to that unit.

File I/O threads. The default context's file I/O pool is max(thread_budget(), 8); decoding stays at thread_budget().

• File I/O threads mostly wait on storage, so they take no core from the job.
• Eight 64³ label chunks written to NFS from one decoding thread took 166–176 ms with one file I/O thread and 83–89 ms with 4 to 16.
• capped_context(file_io=) sizes the file I/O pool separately. Its default is unchanged.

Before and after, one-thread workers (OMP_NUM_THREADS=1), default context:

	0.7.1	0.8.0	zarr-python 2.18
NFS, one 64³ label chunk	23.0 ms	12.2 ms	13.1 ms
NFS, 8-chunk label block	159 ms	66.5 ms	102 ms
NFS, 8-chunk EM block	330 ms	202 ms
Local disk, 8-chunk label block	51.5 ms	4.1 ms	2.9 ms

Create open timed out (cellmap-analyze's question).

• The open in wa.create is now timed out and retried under retries.
• Suppose a retry finds that the array was already made by an earlier attempt of its own, one that timed out. It adopts it after the usual comparison, even under exists="error".
• That open runs on a child context that keeps "os" locking, so create-if-absent stays atomic.

Tests.

• 189 pass locally (1 opt-in stress test skipped). They also pass on tensorstore 0.1.68 with Python 3.10.
• The opt-in lost-wakeup stress test passes: the watched context stranded 0 tasks.
• ruff passes, and mkdocs build --strict is clean.

https://claude.ai/code/session_0141wgF8mkGbjGfBFKfKhp64