[mwarkentin]

Ticket: PRODENG-2971

Description:

Environments affected:
Only processes with SENTRY_REQUEST_RSS_METRICS set (currently the getsentry web canaries in SaaS US, DE, S4S2). Off everywhere else, including self-hosted.

Change plan:

AI Commentary

Why. After ~55 hours on the web canaries, request.rss_growth names the endpoints behind most large RSS jumps (repository.get_repositories, the issue events lists, the GitHub webhook, chunk-upload, ...). Reading the code points to specific causes, but confirming them needs to know how each request was made: e.g. whether the events lists are hit with 100-row pages by API clients, and whether the GitHub webhook spikes are big push events.

What it does. The log line gains these fields (the metric and its tags are unchanged, so no new cardinality):

• duration_ms, status_code, request_bytes (from Content-Length), response_bytes (none for streaming responses)
• user_agent (truncated to 200 characters)
• webhook_event, from X-GitHub-Event, X-Gitlab-Event or X-Event-Key
• organization, the organization_id_or_slug URL argument, to tell whether one organization accounts for the outliers
• param_<name> for an allowlist of query parameters (per_page, limit, full, dataset, sort, expand, collapse, statsPeriod, referrer), plus field_count and query_length. Search text and other parameters are never logged, only the query's length.
• worker_request_number and worker_uptime_s. A new worker's first requests grow RSS while it imports modules and fills caches; these separate warm-up from real offenders.

Keys are flat so each is a queryable attribute in Sentry logs.

Safety. Reporting runs in the middleware's finally, so it is now wrapped: any failure while building the log line is logged with logger.exception and never fails the request or masks the view's exception (e.g. Django raising on a query string with more than 1,000 parameters). The only new per-request cost is two time.monotonic() calls; everything else runs only for requests over the threshold.

Tests. tests/sentry/middleware/test_request_memory.py, 22 tests, adding:

• the new fields on a basic report;
• allowlisted parameters logged, search text and other parameters not;
• the worker request count;
• no response size for a streaming response, and none of status or size when the view raises;
• a failure while reporting doesn't fail the request.

prek and mypy pass on both files.

Generated by Claude Code

---

Generated by Claude Code