Storage, Execution & Recovery Internals

Write-Ahead Logging and Checkpoints

Write-ahead logging decouples transaction durability from random data-page writes: log records describing a change become durable before the changed page, and commit acknowledgment waits for the commit record’s required durability. Checkpoints bound recovery work without necessarily making every page clean.
  • WAL-before-data is the ordering invariant.
  • Commit durability is a log boundary.
  • LSNs order log positions and page progress.
  • Group commit amortizes synchronization.
  • A sharp checkpoint is disruptive.
  • A fuzzy checkpoint records a moving frontier.
WAL and fuzzy-checkpoint ordering
Checkpoint models
ModelConcurrent updatesRecovery knowledgeOperational consequence
Sharp/quiescentStopped or tightly constrainedAll prior pages known durableSimple restart; large pause/write burst
FuzzyContinueCheckpoint plus transaction/dirty-page metadataMore recovery logic; smoother foreground work