Changelog category (leave one):

• CI Fix or Improvement (changelog entry is not required)

Changelog entry (a user-readable short description of the changes that goes into CHANGELOG.md):

...

Description

In 30 days 199 stress jobs (158 PRs, 22 on true master, 12 flavours) failed their post-stress
restart. stress_runner.sh:356 was start_server 10 600 || { echo ...; exit 1; }; that explicit
exit 1 precedes :368 check_logs_for_critical_errors, so on all 199 the S3_ERROR No such key
thrown and Lost forever for SharedMergeTree detectors and dmesg -T > /test_output/dmesg.log
never ran. <Fatal> and signal-9 detection is unaffected: stress_job.py:379-393 greps the same
logs. What is lost is those two data-loss detectors and dmesg.log. Both greps read
clickhouse-server*.log (renamed there at :343), so neither needs a server.

It goes in this caller, not start_server: the function is an end-of-run step, and :356 is the
only such site.

One behavioural consequence, up front. dmesg.log is the kernel-side input to this job's OOM
classifier (stress_job.py:372-377), and :486-488 already turns an OOM-attributed failure into
OK. It is inoperative on this path today, so restoring it will make some currently-red stress jobs
read as OK; its grep matches any container process, not only the server. If they should stay red
instead, that is a small classifier change I can make here; your call.

collect_query_and_trace_logs stays out: it drives clickhouse-local through the same unreachable
object storage, so query_log/trace_log/metric_log stay missing. The three earlier
start_server sites (after 7 start attempts: 10 rows / 1 PR) and upgrade_runner.sh are unchanged.

Validated in the real CI image against the real stress_tests.lib: master emits only the
start-failure row and no dmesg.log; the patch adds both detector rows and the artifact, still
reporting the start failure; deleting the call reverts the result exactly. 3.7 s on a 5 GiB log. In
the reported run the scan would have found nothing; the defect is that it runs on none of the 199.

Related: https://github.com/ClickHouse/ClickHouse/pull/119808
Provenance: Stress test (arm_tsan) on PR #110105 @ 0da1b2081
(report).

Workflow [PR]
Sync PR [sync-upstream/pr/121125]
<!-- CI automatic block end :ci_links: -->