Repository navigation
fetch #11
Description
Activity
I believe we should be much more descriptive about what we want to talk about, and add a Github issue template.
My ideal performance issue would be:
- Add a benchmark
- Mention the specific use case (hot path) for this performance issue
- Provide solutions (if available)
Reacted by Tony Gorez, Ulises Gascón and Carlos FuentesThis is mostly referring to the WHATWG streams or there is something more underneath?
Reacted by Robert NagyI don't think this is done. I haven't seen any evidence of it, I still see a significant gap compared to http.request.
Reacted by Carlos FuentesHere is the latest data I have collected on Node.js v19.14.0:
$ CONNECTIONS=10 PIPELINING=10 PORT=3000 PARALLEL=100 SAMPLES=1000 node benchmarks/benchmark.js │ Tests │ Samples │ Result │ Tolerance │ Difference with slowest │ |─────────────────────|─────────|─────────────────|───────────|─────────────────────────| │ undici - fetch │ 100 │ 314.64 req/sec │ ± 1.29 % │ - │ |─────────────────────|─────────|─────────────────|───────────|─────────────────────────| │ http - keepalive │ 100 │ 600.19 req/sec │ ± 0.92 % │ + 90.75 % │ |─────────────────────|─────────|─────────────────|───────────|─────────────────────────| │ http - no keepalive │ 100 │ 602.57 req/sec │ ± 0.89 % │ + 91.51 % │ |─────────────────────|─────────|─────────────────|───────────|─────────────────────────| │ undici - pipeline │ 100 │ 1653.26 req/sec │ ± 1.81 % │ + 425.44 % │ |─────────────────────|─────────|─────────────────|───────────|─────────────────────────| │ undici - request │ 100 │ 1905.69 req/sec │ ± 1.99 % │ + 505.67 % │ |─────────────────────|─────────|─────────────────|───────────|─────────────────────────| │ undici - stream │ 100 │ 2185.38 req/sec │ ± 2.05 % │ + 594.56 % │ |─────────────────────|─────────|─────────────────|───────────|─────────────────────────| │ undici - dispatch │ 100 │ 2602.14 req/sec │ ± 0.75 % │ + 727.02 % │We could say that the major blocker are WHATWG streams, but I would not consider this "done".
We still need a 2x throughput increase
Reacted by Benjamin GruenbaumThere is a massive improvement in main:
│ Tests │ Samples │ Result │ Tolerance │ Difference with slowest │ |─────────────────────|─────────|─────────────────|───────────|─────────────────────────| │ undici - fetch │ 100 │ 432.22 req/sec │ ± 1.49 % │ - │ |─────────────────────|─────────|─────────────────|───────────|─────────────────────────| │ http - no keepalive │ 100 │ 598.51 req/sec │ ± 0.88 % │ + 38.47 % │ |─────────────────────|─────────|─────────────────|───────────|─────────────────────────| │ http - keepalive │ 100 │ 647.20 req/sec │ ± 0.84 % │ + 49.74 % │ |─────────────────────|─────────|─────────────────|───────────|─────────────────────────| │ undici - pipeline │ 100 │ 1616.99 req/sec │ ± 1.75 % │ + 274.11 % │ |─────────────────────|─────────|─────────────────|───────────|─────────────────────────| │ undici - request │ 100 │ 1900.22 req/sec │ ± 1.67 % │ + 339.64 % │ |─────────────────────|─────────|─────────────────|───────────|─────────────────────────| │ undici - stream │ 100 │ 2170.23 req/sec │ ± 2.06 % │ + 402.11 % │ |─────────────────────|─────────|─────────────────|───────────|─────────────────────────| │ undici - dispatch │ 100 │ 2457.00 req/sec │ ± 1.79 % │ + 468.45 % │We still need a rough +50% increase.
Reacted by Benjamin Gruenbaum, Rafael Gonzaga, Carlos Fuentes, C. T. Lin, Momen and Gary CryeWe can do it! I think I'll have time next week to work on this.
Reacted by Matteo CollinaAccording to my benchmarks (beware that it is flaky), the pull request to add
Adadoubles the performance ofundici - fetchfrom984 req/secto1823 req/sec
Reacted by Toni Villena, Rajesh Khetarpal, Marvin Hagemeister, Carlos Fuentes, Mateo Nunez, Alessandro Commodari, Tony Gorez, Trygve Lie, Momen, Gary Crye and 1 moreReacted by Toni Villena, Gabriel R. Antunes, Marvin Hagemeister, Tony Gorez, Momen, Gary Crye and Vinicius LourençoThat actually should not just improve
fetchbut improve all the methods right? So, in theory, fetch will still be slower thannode-fetch.That actually should not just improve
fetchbut improve all the methods right? So, in theory, fetch will still be slower thannode-fetch.I'm not really quite sure. The amount of
new URLin undici might be greater than the usage innode-fetch.What about the WHATWG, has been any further research for them?
This keep being closed, but either all the fixes needed are not released yet or there is some other problem:
[bench:run] │ Tests │ Samples │ Result │ Tolerance │ Difference with slowest │ [bench:run] |─────────────────────|─────────|──────────────────|───────────|─────────────────────────| [bench:run] │ undici - fetch │ 100 │ 2069.16 req/sec │ ± 1.10 % │ - │ [bench:run] |─────────────────────|─────────|──────────────────|───────────|─────────────────────────| [bench:run] │ http - no keepalive │ 100 │ 3525.25 req/sec │ ± 1.08 % │ + 70.37 % │ [bench:run] |─────────────────────|─────────|──────────────────|───────────|─────────────────────────| [bench:run] │ http - keepalive │ 100 │ 5489.39 req/sec │ ± 2.56 % │ + 165.30 % │ [bench:run] |─────────────────────|─────────|──────────────────|───────────|─────────────────────────| [bench:run] │ undici - pipeline │ 100 │ 6922.14 req/sec │ ± 1.82 % │ + 234.54 % │ [bench:run] |─────────────────────|─────────|──────────────────|───────────|─────────────────────────| [bench:run] │ undici - request │ 100 │ 9877.68 req/sec │ ± 1.47 % │ + 377.38 % │ [bench:run] |─────────────────────|─────────|──────────────────|───────────|─────────────────────────| [bench:run] │ undici - stream │ 100 │ 11218.29 req/sec │ ± 2.20 % │ + 442.17 % │ [bench:run] |─────────────────────|─────────|──────────────────|───────────|─────────────────────────| [bench:run] │ undici - dispatch │ 100 │ 12066.94 req/sec │ ± 2.23 % │ + 483.18 % │ [bench:run]As you can see there is a staggering gap between what is possible (undici
dispatch()) and the cost of running the fetch API.This issue should not be closed until fetch reaches the same level of
http.request()with keepalive enabled.Reacted by xr0masterI believe fetch is now comparable (and faster) than
node-fetch. Closing this as concluded.
Refs: nodejs/undici#1203
Refs: #9