Hello community!
I'm making an internal modificati...
# all-things-deploymenti
important-pager-98358
07/28/2023, 1:14 PMHello community!
I'm making an internal modification to the DataHub chart to enable HPA for standalone consumers, GMS and Frontend. I saw that the GMS can use two caching technologies but one of them is only used when there is more than one replica. Can I use hazelcast even though I only have a single GMS replica?
Is there any other problem that HPA can bring about by autoscaling the components?
Everything going well, I can later make this contribution to the community.
b
brainy-tent-14503
07/28/2023, 6:22 PMNice! Hazelcast can be used with one replica, we just didn’t want to cause confusion as well as deal with the docker-compose settings for quickstart
brainy-tent-14503
07/28/2023, 6:23 PMNote that the parallelization limit for the consumers is equal to the # of partitions in the kafka topics. So scaling replicas of the consumers will not work unless the # replicas <= # of topic partitions
brainy-tent-14503
07/28/2023, 6:24 PMI think particularly for the consumers, the challenge for HPA will be that the operations are typically not memory or cpu bound but i/o bound either in reading or writing.
brainy-tent-14503
07/28/2023, 6:26 PMPlease share what you find out, I think GMS can benefit the most from HPA
i
important-pager-98358
07/28/2023, 7:13 PMThanks for the feedback @brainy-tent-14503. We are thinking for the frontend to do a scaling based on ingress traffic and for gms based on resources. As for consumers, we are evaluating exporting kafka metrics through prometheus and create a custom metric based on topic lag. With this, we should be able to scale consumers at times such as high parallelism of ingestions or execution of index restoration job
important-pager-98358
07/28/2023, 7:15 PMI'm going to develop here internally, and everything working as expected (passing the stress tests) I intend to make a contribution to the community chart.