Report rate-limit calls that silently share or skip a limit
What does this MR do and why?
Gitlab::ApplicationRateLimiter now reports two calls that quietly weaken a limit.
It raises in development and test, and tracks with Gitlab::ErrorTracking in production.
Production behaviour and Redis keys do not change.
| Call | What happens today | After this MR |
|---|---|---|
Scope has no value for any of the rule's characteristics, e.g. { user: nil } on a %i[user] rule |
labkit fills every slot with _unknown_, so all such callers share one counter |
InvalidScopeError in dev/test, tracked in prod |
resource: passed for a key whose rule has no count_distinct |
throttled? returns false, so the limit never applies |
InvalidResourceError in dev/test, tracked in prod, still returns false |
A partly filled scope still passes, e.g. { project: project, user: nil } on a %i[project group user] rule.
"No value" uses the same test as labkit's _unknown_ fallback: nil, an empty string, or an unsaved record.
ci_lint and notes_create keep their shared anonymous counter. Anonymous callers reach both with user: nil: lint on a public project, and comment endpoints that check the limit before the access check. That counter is their only app-level cap, so the check skips these two keys when every value is nil. An empty string or an unsaved record still reports.
Design notes
- Both checks live in
ApplicationRateLimiter, next to the existingvalidate_scope_keys!, which already raises in dev/test and tracks in prod for unknown scope keys. No labkit-ruby change. - When a scope has unknown keys, only that error is reported, so one bad call gives one event.
- The
resource:check sits after the allowlist and zero threshold/interval early returns, so it does not fire for those calls. Any normal call to the same key still reports it. SHARED_ANONYMOUS_COUNTER_KEYSholds the two exempt keys. Moving their anonymous callers to an IP-keyed rule, liketest_report_show_unauthenticated, is a decision for each owner.Labels::DestroyServicehas no anonymous caller, but its EE audit spec passes a nil user. That spec now callsexecute_without_rate_limiting.- The three keys that take
resource:today (ai_users_per_ip,unique_project_downloads_for_application,unique_project_downloads_for_namespace) all havecount_distinct.
References
Resolves https://gitlab.com/gitlab-org/gitlab/-/work_items/637551
How to set up and validate locally
bin/rspec spec/lib/gitlab/application_rate_limiter_spec.rb spec/lib/gitlab/application_rate_limiter/Locally this gives 165 examples, 0 failures (FOSS_ONLY=1). With the old application_rate_limiter.rb, the four nil-scope and resource: raise/track specs fail.
Review checklist
- The all-nil check does not reject a partly filled scope
- Production path tracks and does not raise, for both checks
- No change to Redis key shape or to the return value of
throttled? - Error tracking volume: a tier-3 run found anonymous callers for
ci_lintandnotes_createonly, and both are exempt. Any other all-nil caller would report on every call. - The exemption covers
nilonly, not an empty string or an unsaved record - Teammate review done before maintainer review
MR acceptance checklist
Evaluate this MR against the MR acceptance checklist. It helps you analyze changes to reduce risks in quality, performance, reliability, security, and maintainability.