Operations
How to run a durable Eelgrass database directory: on-disk layout, health signals, compaction, and change feeds. Capability honesty lives on Status.
Which page to read
| Topic | Page | When |
|---|---|---|
What lives under db_dir | Database layout | Opening, backing up, or recreating a DB |
| Trust signals and rewrite | Health and compact | Corruption, WAL lag, eelgrass compact, /ops |
| Live change feed | CDC | Per-tenant tails, LSN resume, Studio live mode |
Wire formats and every route: HTTP API. Commands: CLI. Studio UI: Studio.
Crash expectations
On open the engine reads Meta (superblock + catalog roots) and replays the WAL after the last checkpoint. Committed work that reached a durable WAL sync is re-applied. An atomic that crashed after Begin / Write but before Commit is discarded.
Torn page slots fail checksum and are not treated as good data. See Health and compact. Details of atomic commit vs crash points: Transactions.
When to recreate the DB directory
Recreate (delete or move db_dir, then open a fresh one) when:
- Open fails with an incompatible / legacy Meta error (no in-place migrate).
- You intentionally want a clean engine of record after a bad experiment.
- Health reports corruption and you have no better recovery path than reseeding.
Product export/import packaging is not shipped yet. See Status. Moving the whole directory as a unit is the durable unit of record today.