Closed bikeshedder closed 5 years ago
One useful statistic to know there would be how many postgres connections are actually used in each case. I suspect the async libraries are using far more connections than the blocking r2d2.
As you can see in the code I use the same pool size for all three implementations:
const POOL_MIN_SIZE: u16 = 4;`
const POOL_MAX_SIZE: u16 = 16;`
They were not extracted into constants before. I just changed that to make it more obvious that the pools are configured exactly the same.
I also made sure to run htop
with a filter and could see 48 (3x16) DB connections to PostgreSQL which are all evently utilized, depending on the test currently running.
On my machine, with a release build of your test binary, I get
Running 2m test @ http://localhost:8000/v1.0/event_list_l337
4 threads and 128 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 4.12ms 567.15us 29.99ms 78.09%
Req/Sec 7.79k 164.50 8.97k 73.35%
Latency Distribution
50% 4.04ms
75% 4.37ms
90% 4.77ms
99% 5.98ms
3720626 requests in 2.00m, 1.49GB read
Requests/sec: 30994.28
Transfer/sec: 12.71MB
Running 2m test @ http://localhost:8000/v1.0/event_list_r2d2
4 threads and 128 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 4.65ms 2.41ms 50.40ms 88.78%
Req/Sec 7.23k 316.71 27.38k 98.71%
Latency Distribution
50% 3.69ms
75% 4.42ms
90% 7.49ms
99% 14.98ms
3453682 requests in 2.00m, 1.38GB read
Requests/sec: 28756.74
Transfer/sec: 11.79MB
Running 2m test @ http://localhost:8000/v1.0/event_list_bb8
4 threads and 128 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 4.96ms 322.86us 32.02ms 82.08%
Req/Sec 6.48k 92.65 7.41k 92.40%
Latency Distribution
50% 4.92ms
75% 5.09ms
90% 5.28ms
99% 5.81ms
3096725 requests in 2.00m, 1.24GB read
Requests/sec: 25802.00
Transfer/sec: 10.58MB
which is more along the lines of what I would expect.
I just updated r2d2
to the latest RC which also uses a newer version of (tokio-)postgres
which resulted in a very measurable performance drop for the r2d2 implementation. It might all be related to this: https://github.com/sfackler/rust-postgres/issues/469
It really seams to be an issue of the rust-postgres
crate: https://github.com/sfackler/rust-postgres/issues/469#issuecomment-517764742
I wanted to go full async with the application I'm currently building and tried both bb8 and l337 with tokio-postgres just to find that both performed worse than r2d2 with postgres:
https://bitbucket.org/bikeshedder/actix_web_async_postgres/src/master/
Am I doing anything wrong or is this maybe an issue with tokio-postgres?