Raise Elastic::TimeoutError when Elasticsearch reports timed_out
What does this MR do and why?
Restores the Elasticsearch timed_out check on the Gitlab::Search::Client
path, behind search_client_raise_on_timed_out (default off). With the flag
off the behaviour is byte-identical to today.
Without it, execute_search yields the response straight to its caller, so a
body carrying timed_out: true is consumed as if it were complete.
ResponseMapper#total_count then reads hits.total.value unconditionally — a
number assembled from only the shards that answered — and the user sees a
plausible, smaller, complete-looking result set.
The other client already raises: SearchRequest#execute! in
ee/config/initializers/elastic_client_setup.rb, since 2021. Scopes lost the
check as they migrated, not by any decision to drop it — projects,
milestones, merge_requests and notes still raise; work_items, issues,
epics and every other execute_search caller do not.
References
- Issue: #629244
- Spec:
specs/search-backend-degradation-behaviour/SPEC.mdingitlab-org/search-team/artifacts, merged at433fcacf59f9(acceptance criterion 6)
Screenshots or screen recordings
No user-visible change while the flag is off.
Why raise, test evidence, and what I did not verify
Why raise rather than a partial-results marker
It makes the two client paths agree rather than inventing a third behaviour,
and the affected values surface as a single count where "incomplete" has no
representation. SearchController and the REST search endpoints already render
Elastic::TimeoutError as HTTP 408. GraphQL does not — the exception is absent
from GraphqlController::ERROR_STATUS_MAP — so this ships behind a flag, and
the spec gates full enablement on adding that mapping.
Red-then-green
The fix line reverted, then restored, both runs full:
RED (yield search(es_query)) 11 examples, 9 failures
GREEN (yield search_with_timeout_check(es_query)) 11 examples, 0 failuresThe two examples green in both runs are the controls: flag-off, and a response
with timed_out: false. Files: ee/spec/lib/gitlab/search/client_spec.rb
(chokepoint, including count_only) and
ee/spec/lib/gitlab/elastic/search_results_spec.rb (issues, work_items,
epics, objects and formatted_count for each, plus issue_aggregations).
What I did not verify
The production 200-versus-408 split quoted in the issue above was not
re-measured; this change rests on the source paths, which were re-read, not on
that traffic sample. No live Elasticsearch was available on the machine used,
so the scope-family examples stub the transport adapter rather than running as
:elastic integration specs.