High Pipeline Latency Incident Post-Mortem

Between 15:30 UTC on 8/27 and 14:00 UTC on 8/29 we experienced periods of higher-than-usual pipeline latencies. Between 04:00 and 10:00 UTC on 8/29 most pipelines were completely stopped. At Etleap we want to be transparent about system issues that affect customers, and this post summarizes the timeline of the incident and our team’s response, and what we are doing to prevent a similar incident from happening again.

Number of users with at least one pipeline with higher-than-normal latency.

What happened and what was the impact?

At around 11:30 UTC on 8/27 our ops team was alerted about spikes in two different metrics: CPU of a Zookeeper node and stop-the-world garbage collection (STW GC) time in a Java process responsible for orchestrating certain ETL activities. The two processes were running in different Docker containers on the same host. From this point onwards we saw intermittent spikes in both metrics and periods of downtime of the orchestration process, until the final fix was put in place at 14:00 UTC on 8/29. Additionally, at 15:30 UTC on 8/27 we received the first alert regarding high pipeline latencies. There were intermittent periods of high latency until 10:00 UTC on 8/29.

Incident Response

When our ops team received the first alert they followed our incident response playbook in order to diagnose the problem. It includes checking on potential causes such as spikes in usage, recently deployed changes, and infrastructure component health. The team determined that the issue had to do with the component that sets up source extraction activities, but found no other correlations. Suspecting an external change related to a pipeline source was leading to the increased garbage collection activity, they went on to attempt to narrow down the problem in terms of dimensions such as source, source type, and customer. Etleap uses a Zookeeper cluster for things like interprocess locking and rate limiting, and the theory was that a misbehaving pipeline source was causing the extraction logic to put a significant amount of additional load on the Zookeeper process, while at the same time causing memory pressure within the process itself. However, after an exhaustive search it was determined that the problem could not be attributed to a single source or customer. Also, memory analysis of the Java process with garbage collection issues showed nothing out of the ordinary.

Continue reading “High Pipeline Latency Incident Post-Mortem”