0014. Drift tiers and per-server request budget¶
Status: accepted. Date: 2026-11-14.
Context¶
OpenBao does not notify clients of out-of-band changes, so drift is found by re-reading. With one interval, 30k objects at 10 minutes is about 50 reads per second against one server, and every connection to that server competes without coordination.
Decision¶
Two tiers, both operator flags: access-critical kinds (policies, auth mounts, roles and configs,
audit, groups and memberships) every 15 minutes, everything else every hour, each with ±20% jitter.
spec.driftInterval overrides per object and spec.driftDetection: Disabled turns it off.
OpenBaoServer.spec.rateLimit is one token bucket shared by every connection to that server;
drift reads wait and spec-change reconciles go first. openbao_drift_lag_seconds shows the lag.
Drift reads may go to standbys. The read after the operator's own write is not pinned to a node; point the server address at the active node (see the troubleshooting page) when standbys serve reads.
Consequences¶
- Load on a server is bounded and predictable: objects / interval.
- Drift in standard kinds can stay for up to an hour.
- Tenants share one budget per server; ResourceQuota object counts limit one tenant's share.
- Status: implemented:
--drift-interval-critical(default 15m) and--drift-interval-standard(default 1h).