Ilya Andreev 8995d3aea6 Best-effort tail memory sample and exception-safety hardening for executable UDFs ...

Peak memory: take one best-effort concluding sample when the stdout read reaches
EOF (fired once via a guard in the read buffer), so a peak reached just before
the function closes stdout is captured. This is not continuous coverage of the
post-output reap — it is documented as a single best-effort tail reading. A
short-lived UDF that writes small output and exits before ClickHouse drains
stdout may still report a near-zero peak; this is documented in the event
description and on the sampler.

ShellCommand: capture the reaped child's CPU times by value at reap time instead
of heap-allocating after the child has exited, so a memory-limit `exception` can
no longer fail a query whose child already finished successfully.

UDFProcessSubtreeSampler: the executable path no longer raises the
"PeakMemoryByteSeconds will under-count" degradation flag on a `readPeakRss`
failure — the child's lifetime is not under our control, so a failed read is
usually a legitimately-exited child rather than a sandbox denial. The pool path,
where the borrowed worker is provably alive during sampling, still flags it.

Also correct stale doc comments (the sampler lifecycle bullet and
`recordExecutableFinished` idempotency), clarify the `pollPid` timeout-vs-error
return, and remove the post-EOF exclusion test whose assertion no longer matches
the best-effort tail contract.

2026-06-09 10:25:24 +02:00

..

test_udf_echo.xml

Best-effort tail memory sample and exception-safety hardening for executable UDFs

2026-06-09 10:25:24 +02:00