Distributed Systems Field Notes

Write-Ahead Logs: Durability Without the Fsync Tax

Issue #31 of “Distributed Systems Field Notes” — Part III: Storage Engines

The Quiet Contract Every Commit Signs

Tuesday morning a reader sent me a trace from a payments service that had started losing roughly one write in nine thousand after a rack lost power. Nothing in the application code had changed that quarter, and the disks reported themselves healthy. The culprit was a log that had been told to flush, but never told to wait — a single buffered handoff between the database and the kernel, sitting quietly in the gap where durability is supposed to live.