Skip to content

fetch #11

Closed
Closed
fetch#11
Labels
benchmark-neededAdd to issues that does not have a benchmarkhelp wantedExtra attention is needed

Description

@ronag
ronag
opened on Nov 14, 2022

Refs: nodejs/undici#1203
Refs: #9

Activity

  1. ronag
    added
    performance-agenda
    on Nov 14, 2022

  2. anonrig commented on Nov 17, 2022

    @anonrig
    anonrig
    on Nov 17, 2022
    Member

    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)
  3. anonrig
    added
    benchmark-neededAdd to issues that does not have a benchmark
    on Nov 17, 2022

  4. metcoder95 commented on Nov 29, 2022

    @metcoder95
    metcoder95
    on Nov 29, 2022
    Member

    This is mostly referring to the WHATWG streams or there is something more underneath?

  5. anonrig
    closed this as completedon Jan 23, 2023

  6. mcollina commented on Jan 23, 2023

    @mcollina
    mcollina
    on Jan 23, 2023
    SponsorMember

    I don't think this is done. I haven't seen any evidence of it, I still see a significant gap compared to http.request.

  7. mcollina
    reopened this on Jan 24, 2023

  8. mcollina commented on Jan 24, 2023

    @mcollina
    mcollina
    on Jan 24, 2023
    SponsorMember

    Here 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

  9. RafaelGSS commented on Jan 24, 2023

    @RafaelGSS
    RafaelGSS
    on Jan 24, 2023
    Member

    @mcollina today's release will include a perf improvement on webstream errors. It should improve a bit.

  10. mcollina commented on Jan 24, 2023

    @mcollina
    mcollina
    on Jan 24, 2023
    SponsorMember

    There 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.

  11. RafaelGSS commented on Jan 24, 2023

    @RafaelGSS
    RafaelGSS
    on Jan 24, 2023
    Member

    We can do it! I think I'll have time next week to work on this.

  12. anonrig commented on Jan 30, 2023

    @anonrig
    anonrig
    on Jan 30, 2023
    Member

    According to my benchmarks (beware that it is flaky), the pull request to add Ada doubles the performance of undici - fetch from 984 req/sec to 1823 req/sec

    Screenshot 2023-01-30 at 3 46 11 PM

  13. anonrig commented on Jan 30, 2023

    @anonrig
    anonrig
    on Jan 30, 2023
    Member

    We still need a rough +50% increase.

    @mcollina, this is the rough +100% increase you were looking for (I guess).

  14. RafaelGSS commented on Jan 30, 2023

    @RafaelGSS
    RafaelGSS
    on Jan 30, 2023
    Member

    That actually should not just improve fetch but improve all the methods right? So, in theory, fetch will still be slower than node-fetch.

  15. anonrig commented on Jan 30, 2023

    @anonrig
    anonrig
    on Jan 30, 2023
    Member

    That actually should not just improve fetch but improve all the methods right? So, in theory, fetch will still be slower than node-fetch.

    I'm not really quite sure. The amount of new URL in undici might be greater than the usage in node-fetch.

  16. metcoder95 commented on Jan 30, 2023

    @metcoder95
    metcoder95
    on Jan 30, 2023
    Member

    What about the WHATWG, has been any further research for them?

  17. anonrig
    closed this as completedon Feb 6, 2023
  18. mcollina
    reopened this on Feb 22, 2023

  19. mcollina commented on Feb 22, 2023

    @mcollina
    mcollina
    on Feb 22, 2023
    SponsorMember

    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.

  20. anonrig
    added
    help wantedExtra attention is needed
    and removed
    performance-agenda
    on May 29, 2023

  21. RafaelGSS commented on Aug 15, 2025

    @RafaelGSS
    RafaelGSS
    on Aug 15, 2025
    Member

    I believe fetch is now comparable (and faster) than node-fetch. Closing this as concluded.

  22. RafaelGSS
    closed this as completedon Aug 15, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    benchmark-neededAdd to issues that does not have a benchmarkhelp wantedExtra attention is needed

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions