great, I wonder if logger level should be turned down to :error or similar, but can’t quite make out what the go server logs.
outlog
Asd
Oh, I’ve seen this vid. I am preparing a PR for the Peep library too. It appears that bucket search code there was linear, but not binary. My change will improve the bucket search performance about like 30 times (cause the source has around 240 buckets (why???))
7
josevalim
Creator of Elixir
That’s really cool. I also did a pass to optimize the export itself: Optimize prometheus export by josevalim · Pull Request #24 · rkallos/peep · GitHub 
10
pmjoe
Even if Elixir “only” reaches half of Go’s performance it is already a huge win given everything else that Elixir gives/enables. I’m saying this as a main Go developer. Also, if performance is the ultimate focus, C or Rust should be used instead of Go.
5
dimitarvp
Indeed. Golang is optimal for bigger companies with many devs. It allows for easy onboarding in new codebases.
1
Asd
The company I am working in has recently started to use this Peep library and my colleague made the PR to Anton Putra’s repo to add this benchmark.
Outcome is that this Peep library now has some contributions improving performance and detailed reviews, haha
0
Stefano1990
josevalim
Creator of Elixir
Thanks for sharing. I am a bit surprised indeed the new round did not see any improvement, given my pull request yielded a 20-30% improvement locally. I cannot reproduce the failures from the first benchmark on my machine either, I even used a separate machine to push traffic locally and I get 80k req/s for Elixir vs 110k req/s in Go.
However, it is worth saying that it is 100% expected for Go to be faster than Elixir. After all, we are comparing a static imperative compiled language to a dynamic functional language running on a VM, so a reasonable difference (within the same order of magnitude) is expected. For example, TechEmpower main branch puts Go around 5x faster than Phoenix for single-query runs. I am mostly curious about the failure rate, which is not expected, but as far as I know, there is no way to reproduce the benchmarks and the production logs were not shared with us. So there isn’t much we can do at the moment.
20
raphexion
To me, it would be interesting to continue the run the simulation for both languages. To see how the different applications break. I am impressed how the Elixir application can continue when it is clearly over-worked.
In the second video, part 1, Anton stops the benchmarks at around 60k req/s when it starts becoming super interesting. I would love to see the result if he continues to increase the load until say 200k req/s. So we can see both applications completely break.
2
Exadra37
Wow, this topic is full of good insights. I have some good links and food for though to create some LinkedIn posts (not any-more active on Twitter) :).
For anyone wanting to watch the part 1 and part 2 videos while taking linked notes and creating bookmarks I invite you to watch them at:
1