Concurrent writes to a remote-signing catalog go out unsigned and fail 403

从 pyiceberg/io/fsspec.py 中的 _s3() 开始，然后运行 tests/io/test_fsspec.py 中提到的两个测试。当 shared-client 重新配置时 signer 仍保持安装状态，并且两个并发回归测试都通过时，即表示完成；PR #3783 已经涵盖了所报告的修复。

Apache Iceberg version

main (development) — also reproduced on 0.12.0 and 0.11.1.

Please describe the bug 🐞

_s3() unregisters the S3 request signer and re-registers it on an event emitter that fsspec caches and every thread shares. A request signed in that window goes out with no Authorization header — _s3() has already set config_kwargs["signature_version"] = UNSIGNED, so botocore does not sign in its place — and the store answers 403 AccessDenied.

pyiceberg/io/fsspec.py, in _s3():

fs = S3FileSystem(**s3_fs_kwargs)

for event_name, event_function in register_events.items():
    fs.s3.meta.events.unregister(event_name, unique_id=1925)   # opens the window
    fs.s3.meta.events.register_last(event_name, event_function, unique_id=1925)

FsspecFileIO.get_fs caches per thread and every table gets its own FileIO, so a short workload makes hundreds of these cycles against the one shared emitter. PyIceberg's writer is concurrent by default, so no unusual usage is needed to reach it.

Roughly 3–4% of appends fail against a remote-signing catalog, surfacing as an opaque PermissionError: Access Denied out of s3fs, several frames from its cause. Every unsigned request caught on the wire was a PUT of a manifest during commit.

Measured on Lakekeeper 0.13.1 + MinIO (path-style), 90 writes per run:

configuration	append failures	unsigned on wire
as shipped	2 / 4 / 5	2 / 4 / 5
signer registered once (client still shared)	0	0
PYICEBERG_MAX_WORKERS=1	0	0

The second row is the one that matters: the failures stop while the client is still shared, which separates this from a general concurrency problem. The response code is always AccessDenied and never SignatureDoesNotMatch — what an unsigned request produces, not a mis-signed one.

Suggested fix

Drop the unregister. Botocore's HierarchicalEmitter._register_section returns early for a unique_id it already holds, so re-registering an equivalent signer was already a no-op; the unregister only opens the window.

Reproduction

Two tests in tests/io/test_fsspec.py, no credentials and no network:

• test_s3_leaves_a_signer_installed_while_reconfiguring_a_shared_client — drives _s3() and observes the emitter the instant it unregisters.
• test_the_signer_stays_installed_while_another_thread_reconfigures_s3 — the same window seen from another thread. The window is two adjacent statements, so sampling for it blind is a coin flip (20,000 observations caught it zero times); it is held open by delaying only the re-registration, so the timing is deterministic while the defect is not manufactured.

Both fail on main and pass with the fix.

Already covered by an open PR

#3783 (for #3625) removes this unregister as part of registering the signer on an AioSession, for a different reason — a lazily created client not inheriting handlers, failing InvalidRequest. I applied that PR and ran both tests above against it: both pass, so it fixes this as well.

Worth recording rather than closing silently, because the line is wrong for two independent reasons and #3783's own test does not cover this one. If #3783 lands, these two tests are the regression cover for the concurrency window; happy to raise them against that PR instead if a maintainer prefers.

A second, independent defect in the same file — noted, not proposed here

pyiceberg/io/fsspec.py:160 applies the signing service's headers with add_header, which appends, so every header the service echoes back appears twice — 880 of 1615 sign calls ended with a duplicate of a header named in SignedHeaders. On MinIO this is latent: de-duplicating changed nothing, and it did not contribute to the 403s above. SigV4 combines repeated headers comma-separated, so a stricter store may reject them. Mentioned so it is not lost, not to widen this issue.

Willingness to contribute

I can contribute a fix to resolve this bug independently.