Data & Analytics
Databricks Rewind solves a bad-data window, not every failed stream
Choose Databricks pipeline rewind for known bad data, and a different recovery path for corrupted checkpoints or incompatible code.
November 18, 2026
A stream can be healthy as a process and wrong as a data product. If a known bad input reached downstream tables, Databricks Rewind can restore a Lakeflow Declarative Pipeline to an earlier point and replay from there. That makes the incident window the first question. A corrupted checkpoint or incompatible pipeline change calls for a different recovery plan, as Databricks' stream recovery guide explains.
Find the last trusted point
Identify the first bad source record, the last trusted output, and every downstream table or export that consumed the interval. Record the expected reprocessed volume. Rewind is useful when you can name a safe point before the bad data and can tolerate replay. If the defect predates retained state, the team needs another path.
Check eligibility before an incident
Databricks describes Rewind as a beta capability with preview requirements. Confirm the pipeline's channel, configuration, and supported table types in the current documentation before adding it to a runbook. Run a small rehearsal: introduce a reversible bad record in a test pipeline, rewind, and compare counts and business keys after replay.
Separate state damage from data damage
If the checkpoint itself is damaged or a code change cannot read its prior state, a rewind button is not a universal repair. Follow the recovery guidance for that failure class. Preserve incident evidence before resetting anything, then confirm downstream consumers have not kept a stale copy of the bad interval.
Adovo’s view
Recovery speed comes from classifying the failure before touching state. The right question is not "Can we rewind?" It is "Which state is wrong, and what is the last point we can trust?"
