Live-trip mode

The Managed Service for Apache Flink ("Managed Flink") application wos-stream-intelligence computes stays (detected dwell periods) and per-minute live stats from wos.events.v1 onto wos.derived.v1. Managed Flink bills while RUNNING (~$5.30/day, 2-KPU minimum) so the app is stopped by default and bracketed around each trip.

Read first

• flink/README.md — "How it deploys and runs in production", "Configuration", and Known issues (Confluent credential state of the deployed app).
• tools/README.md — "live-up and live-down" section.

Commands

just live-up      # start, restoring from the latest snapshot; waits for RUNNING
just live-down    # graceful stop WITH savepoint → back to $0/day

AWS credentials for account 458443189947; region defaults to us-east-1 (AWS_REGION honored); app name from WOS_MSF_APP (default wos-stream-intelligence). Both scripts are idempotent — they exit early if already in the target state.

Gotchas

• Credential preflight (added July 2026, PR #44): live-up refuses to start while Confluent credentials are placeholders. Two failure modes, two fixes: placeholder set-after-confluent-signup SASL properties on the deployed app mean the PR #44 fix isn't deployed yet — dispatch backend-deploy (property map gains secret.arn) and flink-deploy (JAR that reads it); a bus secret still holding pending-confluent-api-key means just bootstrap-secrets hasn't run with real credentials. As surveyed July 6, 2026 the deployed app still had the placeholder properties, and keeps them until those deploys are dispatched.
• If the preflight cannot read the secret (missing secretsmanager:GetSecretValue), it warns and continues — the job itself then fails fast at startup if the secret is a placeholder.
• live-down takes a savepoint first, so the next live-up resumes state (distance totals, half-open stay clusters) instead of reprocessing the topic.
• A new JAR (uploaded by a manual flink-deploy workflow dispatch — merging alone ships nothing) takes effect only at the next start.
• No Flink alarms exist — a failing job is silent. Check the MsfLogs CloudWatch group.