Aiohttp switch to orjson

[Reskov]

Switch to orjson since it outperforms ujson https://dev.to/dollardhingra/benchmarking-python-json-serializers-json-vs-ujson-vs-orjson-1o16

pip install aiohttp orjson ujson

from aiohttp.web import Response
from aiohttp import web
import orjson
import ujson

def json_response(payload):
    return Response(
        body=orjson.dumps(payload),
        content_type="application/json",
    )


async def orjson_handler(request):
    return json_response(
        {'message': 'Hello, World!'},
    )

async def ujson_handler(request):
    return web.json_response(
        text=ujson.dumps({'message': 'Hello, World!'}),
        content_type="application/json",
    )

app = web.Application()
app.router.add_get('/orjson', orjson_handler)
app.router.add_get('/ujson', ujson_handler)

if __name__ == '__main__':
    web.run_app(app, port=8080)

Local benchmark

 wrk -t8 -c20 -d20s http://localhost:8080/orjson 
Running 20s test @ http://localhost:8080/orjson
  8 threads and 20 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   622.93us    1.14ms  38.03ms   99.60%
    Req/Sec     3.56k   165.63     3.94k    82.90%
  569116 requests in 20.10s, 93.35MB read
Requests/sec:  28310.04
Transfer/sec:      4.64MB

 wrk -t8 -c20 -d20s http://localhost:8080/ujson
Running 20s test @ http://localhost:8080/ujson
  8 threads and 20 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   805.26us    3.19ms  77.12ms   99.31%
    Req/Sec     3.48k   218.00     3.76k    96.52%
  556324 requests in 20.10s, 99.21MB read
Requests/sec:  27675.67
Transfer/sec:      4.94MB

[Reskov]

Builds fail on /updates, but I'm not sure if switching the json response library can cause this issue. Maybe it is because of the update isn't triggered when randomNumber matching previous, and we need to have something similar.

FrameworkBenchmarks/frameworks/Python/fastapi/app_orm.py

Line 152 in 40b97c5

flag_modified(world, "randomnumber")

also looks like this original code may produce duplicate ids

updates = [(randint(1, 10000), randint(1, 10000)) for _ in range(num_queries)]

[Reskov]

@Dreamsorcerer please take a look.

1. The benchmark performance are unstable locally ./tfb --mode verify --test aiohttp So I can not measure the actual impact of switching from ujson to orjson. But orjson.dumps outperforms ujson on a separate benchmark.

2. The changes of the db is necessary to prevent framework tests flaky

3. psycopg2 library is not used.

[Dreamsorcerer]

Changes are now overlapping my PR, but yeah, looks good to me. Either can be merged and then conflicts fixed.

[Dreamsorcerer]

Note that the places your last commit deviates from my changes though, I already micro-benchmarked.
e.g.

> python3 -m timeit 'from random import randint, sample; a = sample(range(1, 10001), 1000); a.sort()' 'for x in a: y = randint(1, 10000)'
100 loops, best of 5: 2.14 msec per loop
> python3 -m timeit 'from random import randint, sample; a = sample(range(1, 10001), 1000); a.sort()' 'x = tuple(zip(a, sample(range(1, 10001), 1000)))
              for y,z in x: pass'
200 loops, best of 5: 1.79 msec per loop

So, maybe better to revert that commit and just let my PR get merged after this one.