We build in public. Every release, improvement, and fix is recorded here.
You can now expose a /metrics endpoint from the Symbiosis control plane for direct Prometheus scraping. The endpoint covers request count, error rate, p50/p95/p99 latency, and PoP availability. Datadog export continues to work alongside it. See the integrations page for configuration details.
Two new PoPs are now live: Tokyo (ap-northeast-1) and Sydney (ap-southeast-2). Starter and Pro accounts automatically route Asia-Pacific traffic to the nearest node. No configuration change required. Average latency reduction for users in Japan and Australia: 38ms in initial measurements.
The ML prediction model previously retrained every 24 hours. That interval is now 6 hours for Starter and Pro plans, which means the model adapts faster to changing traffic patterns. Early-access pilots with irregular spike schedules saw mean prediction error drop by 22% in the week following this change.
gsb rollback [deploy-id] is now available. It reverts all 12 PoPs to the specified previous deploy in under 30 seconds. If no deploy ID is given, it rolls back to the immediate prior deploy. The last 10 deploys are always available for rollback regardless of plan.
The project dashboard now shows each PoP as a separate row with its own request bar chart, error badge, and latency percentile. Click any PoP row to open a side panel with the last 30 minutes of raw request logs. The top-level summary card remains the default view on load.
Under sustained load above 2K requests per second, routing weight updates were taking up to 8 seconds to propagate to all nodes instead of the expected 1-2 seconds. The root cause was a mutex contention in the weight broadcast path. This is resolved. Propagation is back to under 2 seconds at all tested load levels.
The Datadog metrics export is out of beta and available to all Starter and Pro accounts. Set your API key in project settings and Symbiosis starts shipping request, error, and latency metrics to your Datadog workspace within 60 seconds. Latency histograms are reported as distribution metrics for accurate percentile queries.
Webhooks for deploy events and spike alerts now retry with exponential backoff on non-2xx responses (up to 5 attempts over 30 minutes). Events that exhaust all retries land in a per-project dead letter queue visible in the dashboard. You can replay any DLQ event manually once the target endpoint is healthy.
Subscribe to Dispatch, our engineering newsletter. One issue per release, no noise.
