Core concepts

Persistence

Klyro keeps everything in RAM and snapshots the whole keyspace to a dump file. Saves are atomic, the format is versioned, and older dumps still load.

When a save happens

Trigger	Notes
Graceful shutdown	SHUTDOWN, SIGINT, or SIGTERM — including docker stop
SAVE	Writes the dump immediately and replies OK
Autosave	Every save-interval seconds, but only if something changed

On startup the dump file is loaded if it exists. Killing the process outright, a crash, or power loss loses everything since the last save, which is at most save-interval seconds of writes.

The dump format

The file is length-prefixed, because a value may contain a newline: a KLYRO-DUMP 3 header, then one record per key giving its type and absolute expiry, then each blob as its length followed by exactly that many bytes. Saves are written to <path>.tmp and renamed over the real path, so a crash mid-save cannot corrupt the existing dump.

Indexes are rebuilt, not stored

A memory index writes its configuration and its records, never its inverted index or its vector array. Both are derivable, so they are rebuilt on load. That keeps the dump small and leaves one format to maintain rather than two.

Version 1 and 2 dumps still load, so an existing file survives the upgrade. They are rewritten as version 3 on the next save.

Operating it

klyro

SAVE
+OK

CONFIG GET dbfilename save-interval
1) "dbfilename"     2) "klyro.dump"
3) "save-interval"  4) "60"

# Redirect the next save; the current file is left alone
CONFIG SET dbfilename /data/klyro-2.dump
+OK

SHUTDOWN
+OK

In Docker

The image keeps the dump in the /data volume, so mount one:

terminal

docker run -d -p 7171:7171 -v klyro-data:/data ghcr.io/hitesh-s0lanki/klyro:latest

docker stop sends SIGTERM, and Klyro saves before exiting. Compose allows 30 seconds for that; raise stop_grace_period if a large keyspace needs longer.

Snapshot only

There is no append-only log and no replication in 0.1.1. If losing up to a minute of writes is unacceptable, lower save-interval, and treat the dump as what it is: a periodic snapshot, not a durable transaction log.