KaiBackups and operations
What to back up, health checks, logs and keeping the cache healthy.
Backups
Back up three things together:
- Postgres — everything except files.
KAI_ENCRYPTION_KEY— without it, every stored API key, MCP credential and OAuth token is unreadable after a restore; people would have to enter them again.- The object storage bucket — attachments and large tool outputs.
Redis holds only short-lived state (run event streams, rate limits, caches) and doesn't need backing up.
Health
GET /api/health returns 200 with {"postgres":true,"redis":true} when both
answer. The image's health check uses it.
Logs
Kai writes one JSON line per event to stdout: a line per model request (run, model, stop reason, token counts, time to first token, duration, request id), a line per tool call (server, tool, duration, status), and errors. Secrets and message content are never logged.
The prompt cache
Kai is built so that each model request starts with the previous one, byte for
byte, and providers bill most of a long conversation at the cached rate. The
Usage page shows the cache hit ratio; it should be
high. When a step that should have hit the cache doesn't, Kai logs a
cache_miss warning with the conversation id.
Things that legitimately start fresh: changing a chat's model, turning tools on or off (except on Anthropic with mid-conversation tool changes on), a summary replacing the older part of a chat, and the cache expiring (5 minutes by default, or an hour; see BYOK). Prompt caching has the details.
Restarts
On SIGTERM, Kai stops taking new runs, gives running ones 20 seconds, then
marks the rest interrupted. People see what had arrived and can regenerate.
Runs waiting for someone (an approval) aren't affected.