For more realistic benchmarks to iterate over proto improvements we planned, #14965 and perhaps experimenting with CI use with https://github.com/benchmark-action/github-action-benchmark or so, I propose a refactor of the benchmark we want to persist in our textparse package.

To me we want to persist useful stuff to iterate on, but more importantly, something anyone should run before commiting new changes. For this reasons I killed those isolated cases for series, CT, series+metrics, series+metrics reuse. Instead we only have benchmark for realistic use of parser (as we use in scrape.append) AND opt-in (skipped by default) expfmt comparison. Anyone can comment out certain methods for local testing and comparisons.

Additionally:

• I provided example CLI command
• Fixed cases formatting to match official syntax (you can learn more in https://www.bwplotka.dev/2024/go-microbenchmarks-benchstat/).
• Killed unnecessary sinks. With interfaces and internal state, I don't see any difference if I have sinks or not (that total var).
• I changed operation to be = full scrape, not a single sample. I might miss why we did this per sample, but it feels unnecessary and complex. Each sample is different so perhaps it's simpler to think about ops = scrape (which we care about).
• Removed gzip benchmark, I guess it was for some experiment, but not sure if worth to benchmark always not ours gzip implementation we don't implement.

Would love to know your opinion @beorn7 and @bboreham on this 🤗

Here is an example of what this benchmark enables us to do #14965 (comment)