Troubleshoot — Reconciliation

Ingestion accepted but nothing matches

• Name-worker completions (session_name_results / worker logs).
• Engine received session and published recon-work.
• Worker BigQuery fetch/match errors for the date window.

Session stuck / never finalizes

• Redis counters for remaining windows.
• Incomplete complete-batch push deliveries.
• Partial failures should land in DLQ — use bk-recon-work-dlq-replay rather than blind re-upload.

DLQ growing

• Replay cleans and re-publishes, or finalizes partial on exhaustion (README).
• Do not delete DLQ messages without recording an operational decision.

Exception resolve no-ops

• Scope must match partner + tenant + domain + cluster_type (+ status band).
• reopened selects already-resolved rows; other decisions select pending (see contract).
• Read exception resolution before changing client payloads.

404 on reversals /v1/reversals/background

README may list that path; routers expose /v1/reversals/pipeline/background and /gl-reversals/background. Hit running OpenAPI//docs to confirm.

Cost spikes

• Worker fan-out scales with session size; engine may ref-count min instances.
• Name-worker Gemini volume / batch size (≤150 narrations/msg in README) are common drivers.