URL Parsing Performance

[dajinchu]

Description

Ran a benchmark using the npm create vite-plugin-ssr starter repo on the index page. Used bombardier and node --inspect to generate flamegraph.

Looks like URL parsing is taking up a disproportionate amount of time (~50%) and might be good low hanging fruit for a perf boost.

Noticed parseWithNewUrl is:

1. called ~40 times per request
2. Spending a lot of time in node error handling (see flamegraph)

As a proof of concept I commented out the try section and just let the catch run 100% of the time, and assume there is no origin to return - I'm sure a real solution has more nuance, but this gave ~80% increased req/s and shaves ~40ms off avg latency:

before
Statistics        Avg      Stdev        Max
  Reqs/sec      1484.61     124.81    1829.91
  Latency       83.93ms     7.69ms   183.48ms

after
Statistics        Avg      Stdev        Max
  Reqs/sec      2591.98     309.50    3136.87
  Latency       48.13ms     5.61ms   126.46ms

flamegraph after:

Not sure what the intent is behind the origin handling, so not sure what the real fix is, but curious what can be done.

Thanks!

[brillout]

It's quite suprising that new URL() is that slow. I'm thinking Deno and Bun may have a much faster implementation.

The new URL() usage is only about extracting the origin from the URL (needed for many use cases).

I think getting rid of new URL() and using a custom origin extraction would be not only faster but also more compact (less KBs sent to the browser).

(The couple of days ETA still holds for cumulative: the progress on it has been delayed by a couple of urgent tickets.)

[brillout]

Actually, new URL() is also needed for resolving the pathname which isn't trivial (there are quite a lot of edge cases around escaping for example).