2024-02-12 RabbitMQ message broker partially unavailable

Owned by Joseph (Pepe) Kelly
Last updated: 27 Feb 2024Version comment
2 min read

Date

Feb 12, 2024

Authors

@Joseph (Pepe) Kelly

Status

Done

Summary

The level of use while maintenance was taking place put the broker into a state of protecting in-flight information https://hee-tis.atlassian.net/browse/TIS21-5712

Impact

Applications were unable to fully use the message broker until RabbitMQ stopped synchronising nodes. After that a number of processing changes went through in significant numbers including overnight updates for Revalidation. ESR Applicants and Notifications which would normally have gone on Monday, went on Tuesday. 3 attempts to sign Conditions of Joining failed.

Non-technical Description

The message broker is a service that helps to connect systems without needing them all to be 100% available and responding instantly all of the time.

The broker has several coordinated computers (nodes) to ensure it functions properly. In the weekly maintenance period, there were so many “in-flight” messages that when one of the nodes was replaced and needed to catch up with the other two, it used too much of the memory from those 2 remaining nodes. That placed the broker into a protective “quarantine” mode, allowing messages to be read but restricting the publishing of messages.

We use the message broker for most of how we process information to and from ESR, also for some of how we pull information from the GMC and for recording when Conditions of Joining are signed by trainees.

Trigger

Automated maintenance with large queue

Detection

Alerts in monitoring channel

Resolution

  1. Stopped the synchronisation across nodes in the cluster

  2. Horizontal scaling consumers of queues with lots of message

Timeline

 

  • Feb 11, 2024 22:00-23:59 - Message Broker maintenance window

  • Feb 11, 2024 22:50 - Broker nodes begin being restarted

  • Feb 11, 2024 22:56 - One of the secondary nodes becomes unreachable from the primary

  • Feb 11, 2024 23:28 - Primary and 1 secondary enter an “Memory Alarm”state, although memory reported for the broker did not reflect this

  • Feb 12, 2024 08:41 - First of 3 failed attempts of a trainee to digitally sign conditions of joining

  • Feb 12, 2024 08:45 - Picked up alert and began identifying the issue. We paused some of the feeds in order to free up some

  • Feb 12, 2024 12:11 - Cancelled Sync of largest queue

  • Feb 12, 2024 12:13 - Other queue finished synchronising

  • Feb 18, 2024 - Further maintenance window forced synchronisation, this time with much smaller queues

 

5 Whys (or other analysis of Root Cause)

  • We received “connection timeout” alerts because the message broker was unavailable.

  • The broker was in a “Memory Quarantine State”

  • There was a node started and required synchronising over 1.5M messages

  • Message broker underwent maintenance (restarts/replacement) with too many messages queued up

  • 2 microservices with audit queues had stopped consuming messages.

Action Items

Action Items

Owner

Comments

Action Items

Owner

Comments

Simulate the failure / test pre-emptive alerting

@Joseph (Pepe) Kelly

Mini-hack PoC prep

Rearchitect:

  • Splitting the messaging to limit the impact of spread?

  • SQS instead of RabbitMQ

  • Make neoaudit containers:

    • re-establish connection

    • prefetch / batch consuming

  • PoC to show how flow control would be implemented

 

Mini-hack PoCs & improvements to python scripts:

e.g. time-based scaling of audit consumers (based on when messages are usually published)

Alert on a certain queue size

 

 

 

Lessons Learned

  •  

Slack: https://hee-nhs-tis.slack.com/
Jira issues: https://hee-tis.atlassian.net/issues/?filter=14213