Cache the value of the function port from k8s api

[dannyqiu]

Issue Ref: vmware-archive/kubeless#826

Description:

Address the issue of constantly hitting the k8s api to get the port associated with a function. We only need to retrieve this value once upon creating the consumer process.

TODOs:

• Ready to review
• Automated Tests
• Docs

[andresmgot]

Hi @dannyqiu, thank you for this PR!

I think we are overcomplicating the solution a little bit here. For me a better solution would be to directly remove any call to the K8s API (see vmware-archive/kubeless#826). In my opinion (and we have talked about this in the past) we should remove the possibility of specifying a custom port and use always the port 8080.

For this case I would just add an env var FUNC_PORT that if is set is used instead of calling utils.GetFunctionPort. That way we don't need any complex code nor the dependency you are adding. What do you think?

[dannyqiu]

I agree that a simpler solution can be used. I like the idea of enforcing the port to be 8080, since this is already the default behavior if you don't use a custom deployment.

[wayne-thomson-cxa]

Hi there. I also seem to be experiencing the 5qps limitation reportd in kubeless/kubeless#826). Was this fix abandoned or replaced by something else?

[andresmgot]

we never changed this but I don't think that a limitation of 5 qps is a limitation because of calls to the kubernetes API. See the discussion here: vmware-archive/kubeless#826 (but unfortunately no progress there)

[wayne-thomson-cxa]

we never changed this but I don't think that a limitation of 5 qps is a limitation because of calls to the kubernetes API. See the discussion here: kubeless/kubeless#826 (but unfortunately no progress there)

Thanks @andresmgot. I have read that. Do you have better throughput in clusters you are working with than 5 messages per second or do you experience the same behavior also? I thought from these two threads that wdamron and @dannyqiu identified the k8s API as the issue (hence this issue)?

Whether my topic has 1 or 10 partitions which my producer writes to the consumers seem be be 5/s per controller. As I increase the number of controllers I see a corresponding linear throughput. I just want to check that I am not doing something wrong.

[andresmgot]

You are probably don't. I think this issue was never fixed but I doubt the bottle neck is the K8s API for resolving the port. If you have time to investigate it, you can modify the function getFunctionPort to simply return 8080:

https://github.com/kubeless/kafka-trigger/blob/68a3798a36e725fb15c2dac4d4dc56274d075c3b/pkg/utils/event_sender.go#L40

And check if that helps.