[Elixir/phoenix] Implementing suggestions from @josevalim

[atavistock]

Conversation in #9198 lead to some meaningful optimizations

Specifically

• Obtain one db connection for all queries within a request.
• Optimized flow of getting world ids

[josevalim]

This is beautiful, thank you for working on it! ❤️

The only other feedback I have, which we should explore in a separate pull request anyway, is to increase the pool size. For multiple queries, we are not getting more performance from high concurrency numbers, and I assume this is either because our pool size is too small (we use 50, some of the Rust repos use 1024) or our polling is a bottleneck (it is a single process).

I am away from my computer with 10 cores, so when I am back home, I want to run some benchmarks on top of Ecto polling and try to get some numbers. In particular, I may want to add multiple pools inside Ecto, and then we can try running 8 pools with 32 connections each or similar. But we should probably wait for this PR to be merged and see a trial run, so we can see baseline numbers. :)

If you want, I can keep you posted on the Ecto efforts. Have a fantastic week!

[atavistock]

@srcrip I just wanted to drop a note that this PR was in flight to do Repo.checkout where you were adding Repo.transaction. I tested with checkout, transaction, and checkout+transaction and in these specific performance tests it seems like just the just using Repo.checkout is the best performance presumably because theres very few writes.

[josevalim]

Correct. transaction is checkout+begin+commit, but if you are only doing reads, you don't need begin+commit. :)