Support compatible prometheus service

[n0rad]

To check prometheus server availabiilty, IsOnline method is checking /api/v1/status/flags.

For our platform prometheus itself cannot handle our size and we use some compatible querier services in front of prometheus that can handle queries but do not have this api.

I'm wondering if it's possible to change the availability with a simple query (vector(1) ?) so it will support any prometheus compatible querier.

for info we are using thanos and cortex in addition to prom to do queries.

[stefanprodan]

I'm wondering if it's possible to change the availability with a simple query (vector(1) ?) so it will support any prometheus compatible querier.

Yes that should be a harmless change. Are you willing to implement it?

[mathetake]

FYI, IsOnline is called only for prometheus and during the startup time of Flagger's main goroutine. It is never be called in canary's control loop.

I think this is a little bit confusing since metric analysis phases do not invoke IsOnline and not check the availability of metric server.

@stefanprodan I think we'd better refactor IsOnline's behavior so that this is not prometheus specific, or just remove IsOnline method. Any thoughts?

[stefanprodan]

The IsOnline method is intended for debugging Flagger's install. Before this, we got many issues from users that misconfigured the Prometheus URL, having a sanity check at startup helps users detect that.
What I think we should do is call IsOnline on canary initialisation for all metric providers used in the analysis metrics section and log a warning since Flagger startup check doesn't cover the metrics templates.