Runnev
RU Sign in

Status

System status

All systems operational

Current uptime, from GET /v1/health, loads live. Historical uptime over the last 90 days is shown per component below.

Components

90-day uptime

Ingest, the component that matters most for producers.

99.97% over 90 days. Each bar is one day; amber marks a day with a degradation noted below.

Incident history

Partial ingest degradation

Resolved in 34 minutes

A deploy to the ingest tier rolled out with an undersized connection pool to the metadata store, causing elevated 503 upstream_unavailable rates on publish for roughly 8% of requests between 14:02 and 14:36 UTC. Subscribe delivery was unaffected. We rolled back, restored the pool size, and added a pre-deploy check that fails the rollout if the pool is below the floor for the tier's concurrency.

Elevated subscribe latency in one region

Resolved in 51 minutes

Following the HTTP/3 rollout, subscribers routed through one edge location saw first-byte latency on new subscriptions rise to several seconds during peak, traced to a misconfigured connection-reuse limit at that edge. Existing long-lived subscriptions were not affected. We corrected the limit and added first-byte latency to the per-edge alerting that had not covered the new path.

Buffering regression at the edge

Resolved in 1 hour 12 minutes

An edge configuration change re-enabled response buffering on the /v1 path, so subscribers received batches in clumps seconds apart rather than as published, for about 70 minutes. No data was lost; delivery latency was elevated. We reverted the change, pinned buffering off for /v1 in code review, and added a synthetic check that publishes and measures end-to-end batch latency continuously.

Publish errors during metadata migration

Resolved in 22 minutes

A planned metadata schema migration held a lock longer than expected, causing a subset of stream-creation requests to time out with 500 between 03:11 and 03:33 UTC, during a low-traffic window chosen for the migration. Publishing to existing streams and subscribe delivery were unaffected. Future migrations now run in smaller batched transactions that never hold the lock beyond a bounded window.