AnalysisTemplate and AnalysisRun: CAPA practice questions

Analysis is how Argo Rollouts decides, from metrics, whether a release may continue. It uses three related resources. An AnalysisTemplate is a reusable, namespaced specification of what to measure: one or more metrics, how often to take each measurement and which results count as success or failure. A ClusterAnalysisTemplate is the same thing at cluster scope, available to any Rollout, and is referenced with clusterScope: true. An AnalysisRun is an instantiation of a template for one Rollout revision; like a Job, it eventually completes.

Each metric names a provider and its query. Built-in providers include Prometheus, Datadog, New Relic, CloudWatch, Wavefront, Graphite, InfluxDB, Kayenta and SkyWalking, a web provider calls an HTTP endpoint, a job provider runs a Kubernetes Job and judges it by success or failure, and new integrations are added as plugins. successCondition and failureCondition are expressions over the result, for example result[0] >= 0.95. interval sets the time between measurements and count the number of measurements; a metric without an interval takes a single measurement, and one with an interval but no count keeps measuring until the Rollout stops it. failureLimit sets how many failed measurements are tolerated, consecutiveErrorLimit how many errors in a row are tolerated, four by default, and inconclusiveLimit how many inconclusive results are allowed. initialDelay waits before the first measurement.

An AnalysisRun ends Successful, Failed or Inconclusive, and the Rollout continues, aborts or pauses accordingly. A run is inconclusive when a metric has no conditions at all, or when a value meets neither condition, and it then needs a person to resume or abort the Rollout.

Templates are parameterised with args. The template declares arguments and uses them as {{args.name}} in queries, and the Rollout passes values, either literally or from the Pod template hash and other fields with valueFrom. Templates can also be combined: a Rollout can list several, and the controller merges their metrics and arguments into one run. In a canary, background analysis starts with the update, optionally from a startingStep, and runs alongside the steps, while an inline analysis step blocks progress until its run completes. A blue-green Rollout uses prePromotionAnalysis and postPromotionAnalysis instead. dryRun evaluates a metric without letting it affect the Rollout, which helps when introducing a new query.