Real-Time Analytics Platform · View 15 of 21 · Runtime

Replay, Backfill and Correction

How a defect discovered after the fact is corrected without taking the platform down or letting a consumer read a half-corrected table.

Editable source SVG draw.io All views

Why replay is first-class

Bronze retention is what turns a processing defect into a rerun instead of a data loss event
Replay uses a pinned code version so the correction is reproducible and auditable
Shadow tables mean production consumers never observe an intermediate state

Cut-over discipline

Reconciliation gates the swap: variance above 0.1% discards the run rather than publishing it
The swap is an atomic rename; materialized views rebuild behind it
Consumers receive a restatement notice naming the affected time window

Cost and limits

Replay runs on a separate compute pool so it cannot slow the live pipeline
A full 30-day replay is roughly 105 TB of reprocessing and is a planned activity, not an on-call one
Beyond 90 days the data is in archive tier and rehydration time dominates the runbook

◀ Query and Serving Architecture All views Deployment and Infrastructure ▶