Queues are now up and running normally. For a long time the partitioned state failed to heal, even after several rolling restarts. Ultimately we made the decision to stop and restart the entire cluster, which then took longer than expected due to an infrastructure issue.

We run our queue servers in 'pause minority' mode, which prioritises data integrity and preservation of messages already in the queue over availability to either consumers or new publishers for connections to the cluster in the minority.

We'll be investigating the causes of the partition, why it failed to heal, and the issues experienced in doing a stop-the-world restart, and post any updates here.

The us-east-1 region is continuing to experience a very high rate of errors, so we are now redirecting all traffic away from us-east-1 to us-west-1.

Error rates are quickly returning to normal. There are still some residual issues which we are resolving first before we investigate the root cause of the faults in us-east-1 that triggered the high error rates.

Error rates and availability metrics are now back to normal.

We are currently running with one fewer region than normal. Traffic that would have gone to us-east-1 is now going to us-west-1. We will updated this issue once we have identified the fault and brought the region back online.

Traffic is now being routed to us-east-1 again.

Please see the preliminary incident report on this incident: https://gist.github.com/paddybyers/3e215c0aa0aa143288e4dece6ec16285

Our health check system has reported this issue as resolved.
We will continue to investigate the issue and will update this incident shortly.

Update at 14:03 (UTC): we are routing traffic away from us-east to other datacenters while we investigate the issue

Update at 14:30 (UTC): All regions are now affected by the present issues due to cascading failures. We are very sorry for the inconvenience and are trying to restore normal service as fast as possible.

Update at 16:03 (UTC): All regions are now stable and error rates have dropped dramatically. We will continue to inspect each region manually to ensure any issues are resolved forcibly.

Update at 17:30 (UTC): Error rates have been consistently back to normal in all regions. We believe all operations are back to normal. We will be completing a full post mortem now to understand what caused the global disruption, and importantly what caused the ripple effect from one region to another.

Update at 10:30 + 1 day (UTC): Our focus today is on addressing the immediate issues we identified yesterday that caused the significant increase in error rates and latencies globally. Whilst a post mortem is being prepared and will be published in due course, our priority at present is to resolve the underlying issue and prevent a repeat of the incident.

We are currently rolling out an update globally in all regions and expect this will take a few hours to complete.

We have published a summary of the incident affecting the Ably production cluster 30 October 2018, which includes our preliminary investigation and conclusions.

Please see https://gist.github.com/mattheworiordan/2abab206ee4e4859010da0375bcf4b1d for the report.

As mentioned in the report, we are sorry to all our customers for the disruption this resulted in, and are doing everything we can to ensure we are better equipped to deal with incidents more quickly and without disruption in future.

If you have any questions, please do get in touch with us at https://www.ably.io/contact

Error rates have returned to normal.

We will continue to investigate the root cause of the increased error rates that lasted approximately 15 minutes.

We have completed a post mortem on what happened.

At the time of the issue, we deployed an update to the realtime platform that contained a regression relating to the progressive adoption of load as services come online during scaling operations. This meant that, for a short period of time early in the life of a new server, a small proportion of the channels handled by that server had errors. Externally, this meant that a small fraction of channels were inaccessible in us-east for a period of about 20 minutes.

The bug has since been fixed and we're looking into ways to add more simulated tests over the progressive load feature.

Error rates have returned to normal in all regions apart from eu-central-1.

We are investigating the errors in eu-central-1 now, although the error rate in that region is now very low.

As AWS issues in eu-west-1 and eu-central-1 continue (they have now confirmed networking issues), we are re-routing all traffic for the global cluster away from all EU regions. Please note that any traffic routed to the EU only (EU-only storage options, compliance reasons etc) unfortunately will continue to be routed to these clusters.

We believe the EU regions are now reaching stability and intend to route traffic back to EU in the next 15 minutes once some final testing is complete.

All global traffic is now being routed back to both EU datacenters, and everything appears to be normal. We'll continue to monitor closely now.

Whilst AWS is not yet reporting any issues on the status or personal health dashboards, we believe this issue is caused by a fault in the eu-central-1 and possibly eu-west-1 AWS regions. We are led to believe this as dedicated clusters, isolated from the global traffic in those regions, are also affected.

We will continue to investigate the issue and take action to minimise the impact.

AWS continues to report no issues, yet Twitter confirms the issues are widespread.

We have seen that eu-west-1 continues to exhibit problems, so we are now routing traffic away from eu-west-1 for now.

We are seeing stability return in eu-west-1 and eu-central-1. All eu-west-1 traffic is still being routed to other regions. Once eu-west-1 settles fully we'll redirect traffic back.

We are now investigating any residual issues caused by the partitions and instability to ensure no longer term impact on customer traffic.

Please note this issue is not resolved, but was mistakenly closed. We will continue in issue https://status.ably.io/incidents/562

Error rates have returned back to normal.

We will continue to investigate the root cause and update with a post mortem.

The error rates and latency have returned to normal (after 4 minutes). We believe this was caused by a deploy issue which we are investigating.

Our health check system has reported this issue as resolved.
We will continue to investigate the issue and will update this incident shortly.

The upgrade has been rolled back whilst we investigate the root cause of the issue.

In total, the website was offline for some users for no more than 15 minutes.

The database migrations are now complete and all operations are back to normal.

The new code has been reverted now on all nodes and error rates are back to normal.

We'll be completing a post mortem on this issue and post updates within this incident report.

We have completed our post mortem on what happened:

The underlying cause was that we deployed code that was effectively broken. The specific issue was related to interoperability in a mixed cluster, and arose in a specific configuration that we had not tested sufficiently before release

When we deployed the code to eu-central-1, we saw a very large spike in system load - whose root cause was the bug - but due to the cascade of effects when hugely overloaded we were overwhelmed by logs containing secondary errors so we didn't recognise the root cause initially

The load then triggered autoscaling and recovery of nodes across the cluster which, due to a procedural error in making the deploy, caused other nodes to be restarted with the broken code.

We were able to recover the system by suspending recovery, finding the root cause bug, and reverting affected nodes.

However, moving forwards, we are going to improve our deployment strategy so that we can better test multi-region mixed cluster environments.

We apologize for the inconvenience this caused.

The root cause of this was due to a routine deploy in that region, which triggered an unexpected spike in CPU load across the routing layer for the entire eu-central-1 region. The error rates returned to normal at 6:41 UTC.

We are continuing to investigate why this deploy caused the disruption in this region.

The code that caused this issue was rolled back in all regions across all instances where it had been deployed. See https://status.ably.io/incidents/554 for subsequent incident detailing the rollback process.

We will continue to investigate and publish a full post mortem in due course in incident https://status.ably.io/incidents/554.

Until we understand the underlying issue fully, we are routing traffic away from the eu-west-1 datacenter as error rates are still high.

We are confident we have sufficient information now to understand and fix the root cause of the issue.

Our Cassandra upgrades will now be put on hold until a permanent fix is deployed to the global cluster.

We will shortly be directing traffic back to the eu-west-1 cluster. In the mean time, all traffic in EU is being handled by our datacenter in Germany.

All systems are back to normal in eu-west-1 Ireland as of 22:51.

We have a fix that we will deploy and test in the coming days, before we continue our scheduled Cassandra upgrades.

All traffic is now being routed back to eu-west-1 as normal.