Rust 给 Python Web 带来的性能提升

[Baloneo]

对比 FastAPI Robyn Gin json 序列化返回性能 大概的性能对比如下

wrk -t10 -c 10 -d 20s http://127.0.0.1:xxxx/json

FastAPI (async)

Running 20s test @ http://127.0.0.1:8080/json
  10 threads and 10 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   835.68us  798.76us  29.00ms   94.21%
    Req/Sec     1.32k   513.33     2.51k    49.45%
  262867 requests in 20.10s, 38.10MB read
Requests/sec:  13078.90
Transfer/sec:      1.90MB

Robyn (sync)

Running 20s test @ http://127.0.0.1:8111/json
  10 threads and 10 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   358.51us  690.23us  24.51ms   96.09%
    Req/Sec     3.91k     1.70k    9.80k    76.79%
  781679 requests in 20.10s, 75.29MB read
Requests/sec:  38889.89
Transfer/sec:      3.75MB

Robyn (async)

Running 20s test @ http://127.0.0.1:8111/json
  10 threads and 10 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   517.35us  795.62us  32.63ms   95.72%
    Req/Sec     2.42k     0.87k    6.79k    71.78%
  481576 requests in 20.10s, 46.39MB read
Requests/sec:  23959.48
Transfer/sec:      2.31MB

Gin

Running 20s test @ http://127.0.0.1:8888/json
  10 threads and 10 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   293.92us  688.32us  27.60ms   94.13%
    Req/Sec     6.87k     2.66k   13.32k    58.32%
  1368505 requests in 20.10s, 195.77MB read
Requests/sec:  68090.01
Transfer/sec:      9.74MB

robyn 基准参考 https://github.com/sansyrox/robyn-comparison-benchmarks

Rust+Python Python 的下一波春天?

重新修改了返回 json 的方式
```
from robyn import Robyn, jsonify

app = Robyn(__file__)

@app.get("/json")
async def h(request):
return jsonify({
"message": "Hello, world!"
})

app.start(port=8111)

```
测得
```
Running 20s test @ http://127.0.0.1:8111/query
10 threads and 10 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 562.74us 1.79ms 53.50ms 93.07%
Req/Sec 7.45k 3.76k 16.32k 59.89%
1484168 requests in 20.10s, 159.94MB read
Non-2xx or 3xx responses: 1484168
Requests/sec: 73841.00
Transfer/sec: 7.96MB

```

[wuwukai007]

官方文档就是个简单的 demo ，很怀疑持久性，对比之前的 litestar 文档就很齐全

[rrfeng]

不是，我看着数不是被 gin 秒杀了？

[Baloneo]

@rrfeng 是的

[xgdgsc]

https://discourse.julialang.org/t/julia-can-be-better-at-doing-web-a-benchmark/103300 不如优化下 julia 创造 julia 加 rust 的春天

[TArysiyehua]

这么说吧，我用 python 的那一刻就没考虑性能的问题

[so1n]

没加数据库调用的测试都是耍流氓...而这个例子的接口连 IO 处理都没...

[vicalloy]

如果是 Json 序列化性能，可以做 FastAPI 里使用 orjson 进行加速。
https://fastapi.tiangolo.com/advanced/custom-response/#orjsonresponse
如果不想改动现有代码，可以简单的做个 monkey patch 。

import orjson

json.dumps = lambda obj, *args, **kwargs: orjson.dumps(
obj, option=orjson.OPT_NON_STR_KEYS
).decode("utf-8")
json.loads = lambda obj, *args, **kwargs: orjson.loads(obj)

[Baloneo]

@vicalloy 已经是 orjson 了

[Oxonomy]

序列化带来的提升能说明什么呢🤔

[fakeshadow]

哥们儿你那个附言我笑了，压测 404

[Baloneo]

@fakeshadow 确实是错了 估计是 Robyn 框架并发的 bug

[CodeCodeStudy]

类似 Swoole 用 C++做扩展来给 PHP 做性能提升吗？

[Baloneo]

techempower 的数据库查询参考 https://www.techempower.com/benchmarks/#section=test&runid=473fbd63-e8ff-4082-919e-dd1be6f2a716&test=query

[Baloneo]

@Oxonomy 或许可以将更多的模块用 rust python 绑定重写 用 python 调用

[fakeshadow]

@Baloneo 如果你浏览 tfb 可以参考下 robyn 的分数，横向比较也并没有很快的感觉（我对 py 库了解不多）。https://www.techempower.com/benchmarks/#section=test&runid=074e8a70-d6fb-4f10-82f3-43e57c0965b5&test=plaintext&l=hra0hr-35r

[Baloneo]

@fakeshadow 比 fastapi 还是快很多

[huihuiHK]

@app.get("/json")。 Running 20s test @ http://127.0.0.1:8111/query

[lanlanye]

一般来说，瓶颈都不会在这里……

[SenLief]

我都有 python 了，那性能就不是我考虑的东西。

[shinession]

json 序列化不是用 msgspec 吗? 试了比 orjson 快很多