Bartlomiej Plotka

Benchmarks are looking good! I see some allocs on `NewDesc` as the only "downside": ``` │ baseline.txt │ new.txt │ │ B/op │ B/op vs base │ NewDesc/labels=1-11 376.0 ±...

Questions are: * do we want to do this in short-mid term? * do we want to do this before next release OR 3.0 or just long term?

Kind of works, but some statistical bugs are there, to investigate: 

As per https://github.com/bwplotka/prombenchy/pull/14: CPU is same/similar to srv mode or unchanged agent. Memory is initially much better, but then leaking (for both this PR and unchanged agent, confirming what some...

I'm back from the paternity leave and might time to continue this work.

Still relevant (:

I have to fix tests etc, but initial benchmarks looks promising. There might be more to optimize here, didn't look on this iteration profile yet. ``` benchstat -filter="/fmt:PromProto" gogo=scrape-loop-v1.txt custom=scrape-loop-customdecode.txt...

Current iteration: ``` benchstat -filter="/fmt:PromProto" gogo=scrape-loop-v1.txt custom=scrape-decoder-v2.txt goos: darwin goarch: arm64 pkg: github.com/prometheus/prometheus/scrape cpu: Apple M1 Pro │ gogo │ custom │ │ sec/op │ sec/op vs base │ ScrapeLoopAppend/fmt=PromProto-2...

Getting back to this after holidays...

This should be now good to go 🎉 cc @krajorama @beorn7 @bboreham Results looks solid. There's room for further allocation improvements, but I propose we iterate. Some of those needs...