Trino workers autoscaling

[adwiza]

Hi, we need to scale up and down the Trino workers in Trino cluster, can we use this to to solve this issue, or we need to use an alternative approach, could you please cheet to us how to deal with Trino workers autoscaling?

[sbernauer]

Hi @adwiza, I personally would argue it's the responsibility of the Trino cluster itself to scale the number of workers based on e.g. number of running/queued queries, CPU or some other metric. But I'm also happy about feedback, if you have different opinions.

E.g. at Stackable we will probably utilize Horizontal Pod Autoscaling for that, created stackabletech/trino-operator#532 for that

The community helm-chart of Trino uses the mentioned approach once you enable server.autoscaling.enabled docs

[adwiza]

Hi @sbernauer I haven't mentioned, but that we are using the Trino operator and I guess this solution doesn't fit for us, does it?

[sbernauer]

Sorry, did I get it correctly that you are using https://github.com/stackabletech/trino-operator? In this case you sadly need to wait for stackabletech/trino-operator#532 to have auto-scaling for the workers.

[adwiza]

Yes it is, we are using the https://github.com/stackabletech/trino-operator
productVersion: "428"
stackableVersion: "23.11.0"

[soenkeliebau]

Hi @adwiza
while we do not natively support autoscaling yet, what you could do is just not set the replicas in your clusterdefinitions and deploy your own HorizontalPodAutoscaler.
By removing the "replicas" field you tell our operator to not scale the statefulset and that enables an HorizontalPodAutoscaler to scale it instead.
You would need to find your own metrics to scale on though. The official helm chart could give some inspiration here: https://github.com/trinodb/charts/blob/main/charts/trino/templates/autoscaler.yaml

https://docs.stackable.tech/home/stable/concepts/operations/#_performance

[adwiza]

@soenkeliebau it works, thank you, I've added the HPA configuration with needed behavior that's all

[soenkeliebau]

That is great to hear, thank you for reporting back!

[NickLarsenNZ]

Closing in favour of stackabletech/trino-operator#532