Non-technical Description

ElasticSearch saw a sharp momentary increase in utilization on Prod. TIS could not function properly as a result, as ElasticSearch’s unresponsiveness during the spike made requests timeout.

Trigger

Detection

• Monitoring message on Slack at 13:57 BST reports failed health check on TCS Blue. TIS becomes unusable.

Resolution

• Running a security update on the ElasticSearch cluster restarted the servers.

Timeline

• 16 Jun 2021: 13:51 BST - CloudWatch shows a spike in memory and CPU utilisation

• 16 Jun 2021: 13:57 BST - Slack notification about a FAILING Health Check on TCS Prod

• 16 Jun 2021: 14:00 BST - Identified that TCS’s issue regarded a failing connection to ElasticSearch

• 16 Jun 2021: 14:01 BST - Users noticed being unable to use TIS, as the main screen keeps updating

• 16 Jun 2021: 14:15 BST~ish - A security update’s been run as a way to restart the servers (as they clusters can’t be restarted manually)

• 16 Jun 2021: 14:17 BST - Slack notification about a SUCCESSFUL Health Check on TCS Prod

Root Cause(s)

• Memory usage (JVM pressure) increased steadily during the morning, reaching 75% at 12:30 UTC:
  

  

• Once JVM memory pressure reached 75%, then Amazon ES triggered the Concurrent Mark Sweep (CMS) garbage collector. Some memory was reclaimed, but JVM memory pressure again reached 75% at 13:05 UTC, triggering another garbage collection. Garbage collection is a CPU-intensive process, pushing CPU utilisation to 100% between 12:50 - 13:15 UTC.

• It is possible that during this period the cluster was encountering ClusterBlockException and/or JVM OutOfMemoryError; there were definitely cluster performance issues (as per https://aws.amazon.com/premiumsupport/knowledge-center/high-jvm-memory-pressure-elasticsearch/). Error logging has now been enabled on the cluster to provide this level of detail in future.

• There are a range of possible reasons for the steady increase in JVM memory pressure. In a general sense, the cluster may be configured sub-optimally. In particular, the number of shards (5) may be too high for the persons / masterdoctorindex indices, given these both comprise less than 300mb total size. AWS recommends shard size between 10–50 GiB as “too many small shards can cause performance issues and out of memory errors”. Elastic recommends “average shard size between at least a few GB and a few tens of GB”. Benchmarking performance with fewer shards could confirm whether redesigning the cluster would be advantageous.

• During normal usage, the ElasticSearch cluster shows gradual increases in JVM memory pressure, followed by garbage reclamation, in the normal saw-tooth pattern. As JVM memory pressure is a measure of the fill rate of the old generation pool, this reflects the accumulation of long-lived objects (e.g. cached searches?) in memory, and is not inherently problematic.

• Nightly sync runs cause a spike in indexing activity, and JVM memory pressure commonly increases at that point; however, this incident occurred outside of that timeframe. There were no apparent significant sync jobs being run at the time of cluster failure. Given that JVM memory pressure had been steadily increasing during the course of the morning, it is more likely that slightly higher than normal usage, or unusually diverse searches, were causing an accumulation of long-lived objects. (Would Google Analytics confirm this?)
  

Action Items

Lessons Learned