Defer eviction of checkpoint blocks to reduce post-checkpoint cold reads

[krleonid]

Summary

• During checkpoint, newly-written persistent blocks are pinned in the buffer pool instead of being immediately added to the eviction queue
• This prevents the checkpoint process from cannibalizing its own freshly-written blocks when allocating memory for subsequent writes
• After checkpoint completes, all pins are released and blocks become eligible for normal LRU eviction

Problem

Each ConvertToPersistent call writes a block to disk, then adds it to the eviction queue. The next block write needs a buffer, so the pool evicts the oldest queued block — which is the one just written. This creates a cycle where checkpoint evicts its own output, forcing the first post-checkpoint query to re-read everything from disk.

Measured improvement

67MB dataset, 80MB memory limit:

• Before: 23 blocks re-read from disk after checkpoint
• After: 5 blocks re-read (78% reduction)

When data fully fits in memory, post-checkpoint cold reads drop to zero.

Test plan

• Unit tests pass (build/reldebug/test/unittest)
• Verify with SET memory_limit + checkpoint + query pattern
• Verify no OOM when dataset exceeds memory limit (deferred pins are bounded by available memory)

🤖 Generated with Claude Code