sqlx/sqlx-bench/results/2020-07-01-bench_pgpool_acquire/REPORT.md

2020/07/01: pg_pool benchmark added to test pool fairness changes

• Commit: 7c8ef602661c9cb62c8cf43aaf0f8faaf5b0aed5
• Machine specs:
  • (lscpu) Intel(R) Core(TM) i9-9900K CPU @ 3.60GHz
  • 64 GB DDR4 RAM
  • (Distro / uname -a) Arch Linux / kernel 5.6.15-zen2-1-zen
  • (rustc -V) rustc 1.43.0-nightly (834bc5650 2020-02-24)
• Command: cargo bench --features runtime-tokio,postgres -- --measurement-time 30

Connections	Unfair Pool	Fair Pool
5
10
50
100
500
1000
5000

When the pool is set to be fair it appears the average time for acquire() at high contention (500+ tasks using 50 pool connections) actually goes down by up to an order of magnitude since tasks in the waiter queue aren't getting preempted.

This appears to be the solution to our problem of tasks timing out waiting for a connection at high load. At very low contention (5-10 tasks using 50 pool connections), the average time for acqure() increases by perhaps 10% but a difference between 80µs and 90µs is hardly noticeable in real applications.

A normal MPMC channel may care about that kind of overhead, though, which is why most implementations aren't perfectly fair.