Skip to content

Iceberg — Middle

At middle level, focus on this question:

How does querying an old snapshot reconstruct a past table state without needing to read any of the table's current data?

Prerequisite: junior.md.

Every snapshot is a fully independent, self-contained pointer

flowchart LR Snapshot1["Snapshot @ v5\n(Monday)"] --> ML1["Manifest List v5"] Snapshot2["Snapshot @ v8\n(current, Thursday)"] --> ML2["Manifest List v8"] ML1 --> Files1["Data files as of\nMonday (some may have\nsince been superseded,\nbut STILL EXIST on\nobject storage)"] ML2 --> Files2["Data files as of\nThursday (current)"]
-- Query the table as it existed at a specific past snapshot
SELECT * FROM my_table VERSION AS OF 5;

-- Or by timestamp
SELECT * FROM my_table TIMESTAMP AS OF '2024-01-15 00:00:00';

Because each snapshot points to its own manifest list (which points to the specific data files that were part of the table at that version), querying an old snapshot simply means starting the read path from that older snapshot's manifest list instead of the current one — no replay of intervening operations needed at all (unlike Delta Lake's log, where reconstructing an old version conceptually means replaying up to that point). As long as the underlying data files an old snapshot references haven't been physically deleted (via VACUUM/expiration, analogous to Delta Lake's VACUUM), that old snapshot remains fully, independently queryable.

🎓 Takeaway: time travel in Iceberg is a direct consequence of the tree structure from junior.md — an old snapshot is simply a different, independent root pointing into a (possibly overlapping) set of data files, not a state you have to reconstruct by replaying history. This is structurally different from Delta Lake's approach, even though both systems offer the same user-facing "time travel" capability.

Test yourself

  1. Why doesn't querying an old Iceberg snapshot require replaying any intervening commits, unlike (conceptually) Delta Lake's log-based approach?
  2. What would make an old snapshot's data become unqueryable, even though the snapshot metadata itself might still exist?
  3. Why can two different snapshots (an old one and the current one) safely reference some of the same underlying data files, without any conflict?

Continue to senior.md.