Description

• Matrix workers (Kubernetes setup) do not free up RAM memory and tend to always grow bigger (over a few days).
• Kubernetes can not scale down corresponding worker deployments if memory consumption increases continually (we use HPA on memory + cpu)
• Unnecessary worker pods consume database access slots

Steps to reproduce

• run Matrix with workers under kubernetes and use it for a long enough period

Especially with workers handling sync, room and userdir, as these workers tend to grow bigger over time, their HPA tend to create unnecessary new worker pods as their RAM increases.

Version information

• Homeserver: synapse 1.46 (and earlier versions)

• Install method: Kubernetes Deployment

For now, to prevent this we either over allocate memory resources or periodically delete workers biggest pod instances.

A mechanism freeing up unused memory on workers should help Kubernetes properly do its job scaling workers pods up/down.