Previous incidents

Runs on large-2x machines were processing more slowly due to incredibly high volume. One of our customers caused an infinite loop in their large machine tasks which kept triggering more runs. Runs on small and medium machines were unlikely to have been impacted by this.

This resolved shortly after a problematic control plane rollout.

Service in us-east-1 has fully recovered

Larger machines are now dequeuing faster.

We are going to change the packing algorithm we use which means this will be much less likely to happen in the future. That work will begin this week.

The database load is back to normal.

We're still investigating why this happened, the top theory at the moment is a an auto-vacuuming issue possibly to do with transaction wraparound.

This issue was caused by our us-nyc-3 cloud provider taking an abnormally long time to spin up new servers and capacity issues, along with some runs getting stuck after restoring from a snapshot and not completing in under 2 minutes, which also mostly happened in us-nyc-3.

We were seeing crashes on multiple servers in the EU only. It was related to an out of memory issue in our “supervisor” which meant some servers weren’t dequeuing consistently. We’ve changed some settings to allow for more memory. We're still investigating why these supervisor processes were using up too much memory and crashing, and we're monitoring the situation. us-east-1 runs have not been impacted.

This is now resolved.

We're still fully investigating the root cause but there was too much activity with etcd and clean up not happening fast enough. This caused downstream impacts on pods piling up and Kubernetes API rate limits being hit. There were Out Of Memory issues because of this.

Our older Digital Ocean cluster is doing far higher volume that this so we need to determine what configuration is different and making appropriate changes to prevent this happening again.

v3 dequeue times have been greatly improved and we're now seeing sub-500ms p50 dequeue times (it was previously 10-20s p50 dequeue times). We ended up backporting all of the dequeue performance improvements we made in v4 and even made some improvements that we'll be bringing to v4 soon as well.

Queue times are back to normal. We had an unprecedented number of new v4 runs. We've adjusted our autoscaling rules across multiple services to account for this. We are looking into how to avoid this happening as v4 scales up.

This was resolved at 00:29am. Logs and the API started recovering once a valid partition was in place.