Alert investigation — sessions
Load this file when investigating a session-based alert (session-error rate, user-facing failure, session-quality anomaly).
Before you look for a cause
- Replay the firing with the
replayarguments fromget-alert-firing-history, then withcompareTo: "1w". Session counts follow traffic; a rise that matches last week's shape is not an incident.
Investigation shape
- Pull matching sessions via
search-sessionswith the alert's query (fromget-alertswithalertId) and the firing window. Start small — 5–10 sessions sampled from the window. - For each session, walk the timeline via
search-timeline-events. Find the last successful interaction and the first failing one. The transition is the key moment. - Cross-reference errors. If sessions have
has_errors=true, pull the error group for each. If sessions share an error group, that's the common root. - Correlate backend and frontend. Session events often carry trace IDs — follow them to backend traces and logs for the full picture.
- Check for a common trigger. Is it a specific browser? Geography? Time of day? Shared identifier attribute?
What goes in the diagnosis
- What triggered — alert name, threshold, the query pattern matching the affected sessions.
- Likely cause — the common failure point across sessions with one quoted session ID as example. Cite specific timeline events.
- Scope — number of affected users, geographic or demographic clustering, time-window pattern.
- Next steps — concrete: "roll back the <feature> deploy", "disable flag <key>", "check for a degraded third-party script on the target page".
Wrap session IDs
Whenever you reference session IDs in the diagnosis, wrap them in <session-id> tags:
<session-id>c6wX2TtznM6F0QByEiRBoQBFw1Bv</session-id>The UI renders these as clickable links.
Common mistakes
- Skipping the timeline walk and inferring behavior from session attributes alone.
- Paraphrasing errors instead of pulling the error group record.
- Forgetting to wrap session IDs in
<session-id>tags. The UI can't link them otherwise.