# Replay and Recovery

This page explains how ToriiDB loads a database from its snapshot and logs, falls back to an older generation, and how to back the data up.

## Replay

A database is loaded on first use (`sync.Once`):

1. Remove leftover `*.tmp` files.
2. List snapshot numbers from newest to oldest; with no snapshot, start from generation `0` (logs only).
3. For generation `N`, replay `N.snap` (if `N > 0`), then `N+1.log`, `N+2.log`, ... while they exist.
4. On an I/O error, remember it as `loadErr` and try the next older snapshot.
5. On success, clear `loadErr` and keep appending to the last existing log.

Lines are read with `bufio.Reader.ReadBytes('\n')`, so line length is unlimited. `bufio.Scanner` is not used because its 64 KB default line limit made a 70 KB value hide the entire database. Empty or malformed JSON lines are skipped; replay only fails on real read errors, such as a `.snap` path that is a directory.

Replay rebuilds `SET` entries with their expiry and vector, applies `DEL`/`EXPIRE`/`EXPIREAT`/`PERSIST`, and warms the parsed cache of JSON values.

## Backup and Recovery

- Back up the whole data directory, including every `db_N/` directory.
- Close the store (or stop writes) before copying if you need a consistent point-in-time copy.
- When the newest snapshot cannot be read, the database loads from the previous generation automatically.
- When no generation can be replayed, the process keeps accepting appends but refuses to compact, so no snapshot is overwritten.

## Related Pages

- [Files and Writes](/persistence)
- [Compaction and GC](/persistence-compaction)
- [Removed APIs](/removed-api)