Skip to content

Root cause analysis & correlations

Two related screens under Intelligence try to shorten the distance between “something broke” and “here’s why”: one suggests a probable cause once an incident is resolved, the other looks for patterns across multiple services that might be the same underlying problem showing up more than once.

Intelligence → Root Cause Analysis (/gravity/operations/rca) lists RCA candidates - a probable cause, supporting evidence, and (where applicable) a suggested fix, generated after an incident is resolved.

Each candidate shows:

  • A probable cause headline
  • A confidence score as a percentage - colored green above 70%, amber between 40% and 70%, red below that
  • Evidence - an expandable list of the specific signals behind the suggestion (for example, a correlated change request or a cost anomaly around the same time), each with a short description and source
  • A suggested fix, when one was generated
  • Which model produced it, and when

You can filter by status or by a specific incident’s ID - arriving here from an incident’s own page pre-fills that filter for you.

For a candidate still pending review, you can:

  • Accept it - records that your team confirmed it as the (or a) real cause
  • Reject it - records that it wasn’t

Either action is a record-keeping step: it updates the candidate’s status and is logged, but it doesn’t automatically create a runbook, a change request, or any other follow-up action on its own - you still decide and trigger whatever comes next yourself. Both actions require the Member role or higher.

Where a candidate was generated by a language model rather than a simpler rule-based check, the model name is shown alongside its confidence score, so you can weigh a suggestion accordingly - confidence from a rule-based fallback tends to be more conservative than one from a full analysis.

Intelligence → Correlations (/gravity/operations/correlations) surfaces patterns detected across multiple services or signals - for example, a cascade failure, an availability, performance, security, or cost-shaped pattern - rather than a single incident’s cause.

Each finding shows its pattern description, a confidence level (high, medium, or low), a status (new, investigating, resolved, or false positive), the affected services, an evidence summary, and a count of the underlying events behind it. For a new or investigating finding, you can Dismiss it.

Finding typeWhat it’s flagging
Cascade failureOne failure that appears to have triggered others downstream
AvailabilityA pattern in uptime/availability signals across services
PerformanceA shared slowdown pattern across more than one service
SecurityA security-relevant pattern spanning multiple signals
CostA cost pattern correlated with other operational signals