Previous incidents

All but a very small slice of our A40 hardware is back online, and Replicate workloads are processing normally. We again thank you for your patience.

Workloads across all regions are now running normally. We apologise for the disruption, and will working to better improve our ability to shift load between providers in situations like this one.

We identified excessive load on our database. Shortly after the root cause was isolated, our engineering team disabled the problematic code paths.

Degradation of database responsiveness resulted in the general service outage beginning at approximately 6:23UTC.

At this time API, Inference, and Web are fully functional. Any predictions that resulted in errors or were marked as failed during this window can be safely re-run.

We are seeing HF Hub return to full functionality. The backlog of models blocked on interacting (downloading or otherwise) with HuggingFace have recovered.

From approximately 13:05 to 13:11 UTC, our prediction and training creation endpoints were unavailable. Existing predictions and trainings were unaffected, but no new predictions or trainings could be created.

The problem has since been resolved.

This maintenance has concluded.

At this time all of our accelerated content delivery has been restored to full working order. There may continue to be some minor slowness loading weights as the systems complete their operational warmup. We do not expect any significant issues.

This issue has been resolved and predictions are now flowing again.

The problems resolved automatically at 09:41 UTC. We are monitoring the situation.

The models stuck in booting have been fixed, and inference and trainings are both running normally again.

Well, lots of things went wrong today. We've identified what we think are the last few things that were broken and fixed them. A newly rolled out internal queueing service didn't allow traffic from model pods, and that caused our prediction throughput to be far lower than normal, which was impeding recovery.

For an incident of this magnitude we fully understand that many of our customers will want to know what happened. We're starting to piece it together and we'll have a proper write-up f...

We've not seen any failures after 22:50 UTC, so we're calling this incident resolved.

Our investigation revealed that internal DNS lookup failures put a storage cache subsystem into a broken state. Next week we'll be looking into how to make our systems more robust in situations like this one.

Thank you for your patience.

Mitigations are in place and now models are again downloading weights as expected.

All model startups within the affected region have returned to normal.

We've resolved this issue and have re-enabled all affected versions.

The backlog of models has been cleared. Model start time is back to expected times.

A fix has been rolled out to the majority of models and errors rate has returned to normal levels. We will continue to monitor to address any more occurrences of the errors.

Predictions affected by this incident (many on T4 gpus, CPU, and a subset of a100s) will appear to be stuck in the starting phase for an extended period of time. These predictions can safely be cancelled and reattempted.

There have been a number of automated reports the Replicate website has gone down/returned to service. We are investigating the automated systems but do not see any current outages outside of the automated tooling.

The incident is in process of clearing and model start times have returned to normal.

At this time all models in the backlog have finished startup. We will continue to monitor the situation closely.

We are aware of an event in one of our regions that resulted in extended boot times of many models. At this time the incident has resolved. We are actively researching the root cause and will work to build remediations to limit impact of future such events.