Lesson 44 · Checkpointing - Build a SQL Database

Checkpointing

A log that only ever grows makes every open slower and every crash take longer to recover. Today you add a checkpoint - fold the log into a fresh atomic snapshot and truncate it - so recovery time stays bounded.

The goal

Checkpoint by writing a fresh atomic snapshot of the current state and truncating the WAL to empty, without changing what a later open recovers.

Start here - the target

TO DO

Scenario: Folding the log into a new snapshot

Givenan open store whose snapshot "db.txt" holds [1, "alice"] and whose log "db.wal" holds an INSERT of [2, "bob"]

Whena checkpoint runs

Then"db.txt" is atomically rewritten to hold both rows [1, "alice"] and [2, "bob"], and "db.wal" is truncated to zero length

Andreopening after the checkpoint yields exactly those two rows with an empty log - bob is applied once (from the new snapshot), never twice

Background

Replay works, but the log grows without limit - a store that runs for a month would replay a month of mutations on every open. A checkpoint collapses that history: write a fresh atomic snapshot that already includes every logged mutation, then truncate the log back to empty. Recovery after a checkpoint starts from a snapshot that is fully up to date, so there is nothing left to replay until the next mutation.

The ordering is a small but critical durability rule: take the new snapshot first, and only truncate the log after the atomic rename has succeeded. Do it the other way and a crash between truncate and snapshot would lose the logged mutations forever. Done in the right order, a crash at any point leaves either the old snapshot plus the full log, or the new snapshot plus an empty log

both of which recover to the identical state, with each mutation applied exactly once.

Make it work

// Checkpoint():
//   save the current in-memory state via the atomic snapshot (lesson 41)
//   only after the rename succeeds, truncate db.wal to 0 bytes
// ordering matters: snapshot first, THEN clear the log

CheckpointDONE

The store can fold its log into a fresh snapshot and reset the log, bounding both file size and recovery time. Commit and stop here.