Support Endpoint Stats

[arkodg]

Description:

Describe the desired behavior, what scenario it enables and how it
would be used.

Add support for enabling per endpoint stats in envoy
envoyproxy/envoy#29709

https://gateway.envoyproxy.io/latest/api/extension_types/#proxymetrics

Will help with traffic splitting telemetry #2809

[ShyunnY]

I want to try doing this, but I'm not sure if I have the right idea. Do we just need a bool as a switch to enable per endpoint stats on envoy? can you tell me more :)

[arkodg]

yup bool pointer is the way to go here :)

[ShyunnY]

When implementing this feature: I will put EnableEndpointStats in ir.Metrics, since adding this feature will require setting the cluster in the buildXdsCluster function.

There is currently a problem: buildXdsCluster is called by different processXXXListenerXdsTranslation, and the parameters of these functions are not consistent. If we plan to support endpoint metrics for all protocols, this will change the parameters of many functions. Big changes to the code.

Currently we are considering whether to enable EnableEndpointStats only under HTTP, or whether to enable it under all protocols. I want to hear your opinions. Thanks :)

[ShyunnY]

cc @zirain @Xunzhuo

[ShyunnY]

@arkodg WDYT comment

[arkodg]

not sure I exactly understand @ShyunnY, here's what Im thinking

• add a new field under ir.Metrics as youve suggested
• add another field in xds cluster args

  gateway/internal/xds/translator/cluster.go

  Line 40 in a5bedbc

  type xdsClusterArgs struct {

• build/translate in

  gateway/internal/xds/translator/cluster.go

  Line 77 in a5bedbc

  func buildXdsCluster(args *xdsClusterArgs) *clusterv3.Cluster {

[ShyunnY]

Yes, you're absolutely right, that's exactly what I was thinking.

But another problem arises at this time:
The upper caller of buildXdsCluster is addXdsCluster, and addXdsCluster is called by different callers. We will find a phenomenon: *ir.Metrics seems to only appear in processHTTPListenerXdsTranslation:

gateway/internal/xds/translator/translator.go

Lines 118 to 124 in a5bedbc

	func (t *Translator) processHTTPListenerXdsTranslation(
	tCtx *types.ResourceVersionTable,
	httpListeners []*ir.HTTPListener,
	accessLog *ir.AccessLog,
	tracing *ir.Tracing,
	metrics *ir.Metrics,
	) error {

Based on the information above, going back to the original comment, do we currently only support endpoint stats under HTTP? This question has made me hesitant about how to proceed with writing this PR, because I'm not sure.
@arkodg

[arkodg]

we should support for all, the existing field in that IR (virtualHostStats) is related to HTTP, so the initial commit was only applied to HTTP

[zirain]

will this cause high cost of memory?