Backup & recovery
Durability, crash recovery via WAL replay, and data-directory snapshots.
Durability
Writes are made durable through a write-ahead log (one durable segment log per
table-shard). The client is acknowledged at the durable group-commit watermark.
Storage policies choose the durability posture (durability setting), from
synchronous fsync-before-ack to async.
Crash recovery
On startup the engine reconstructs its catalog and in-memory state from the data directory: it replays the DDL and DML journals and rebuilds current-state indexes. Sequence numbers make replay deterministic and underpin exactly-once semantics. Per-part secondary indexes are recovered with their parts (from the CRC'd footer region), not rebuilt from scratch.
Snapshots
The engine's on-disk state lives entirely under --data-dir
(/var/lib/nyxdb in the container): journals plus immutable columnar parts.
TODO-verify: a supported online-backup / snapshot procedure (e.g. filesystem
snapshot of the data directory vs. an engine-native backup command) is not yet
documented for a specific release. Treat --data-dir as the unit to snapshot,
and confirm consistency requirements (quiesce vs. hot copy) for your version.