DataYatriSpark, seen under the hood
Simulate
databricks/lakehouse
intermediate
Connecting…

Time travel: reading the commit log backwards

Read an old version from the commit log, RESTORE as a new forward commit, and the VACUUM that ends time travel.

Lesson 6 of 14 · Databricks path

Explain it at my level

  1. Commit log
  2. Copy-on-write
  3. AS OF
  4. RESTORE
  5. VACUUM
Watch the canvas:_delta_log commitlive data filetombstoned / deletedtime-travel readrestorevacuumLive simulation
_delta_log — ordered commitsv0 · CREATEadd part-000v1 · INSERTadd part-001v2 · UPDATEremove + addv3 · RESTOREfile list of v1Parquet data filespart-000.parquetrows 1–3referenced by every versionpart-001.parquetrows 4–6 · status='pending'livepart-002.parquetrows 4–6 · status='shipped'written by the UPDATETime-travel readVERSION AS OF 1or TIMESTAMP AS OF '2026-09-01'read-only queryReplay log 0 → 1rebuild that snapshotmetadata only · no copyScan part-000 + part-001part-002 is not in the listcurrent version is still 2Version N = replay commits 0..N. The folder alone is not the table.
Versions are replayed commits. AS OF is a read, RESTORE is a commit, VACUUM is the point of no return.