Possible memory leak in Fetch API

[kamdubiel]

Link to the code that reproduces this issue

https://github.com/kamdubiel/next-fetch

To Reproduce

I created simple reproduction with docker compose. There's simple Express API (just to return random JSON data), K6 for stress test and NextJS 14.2.5 with simple dynamic route with async server component.

1. Clone the repository: https://github.com/kamdubiel/next-fetch
2. Run the containers: docker compose up -d
3. Attach to Next container stats: docker stats next
4. Run K6 test: docker compose run k6 run /scripts/ewoks.js
5. Wait for the K6 test to finish - there will be high memory usage even with idling Next server. It will never go down, and in real world app it's causing out of memory exceptions every few days.

Current vs. Expected behavior

Current:
When container starts, it's memory usage is around ~45MB.
After stress test, the memory usage goes up to ~400MB, and stays there forever. Running more tests will increase the idle memory usage.

Expected:
After stress test, when Next is idling, the memory used should drop.

Provide environment information

From Docker container:

Operating System:
  Platform: linux
  Arch: x64
  Version: #39-Ubuntu SMP PREEMPT_DYNAMIC Fri Jul  5 21:49:14 UTC 2024
  Available memory (MB): 63434
  Available CPU cores: 16
Binaries:
  Node: 20.16.0
  npm: 10.8.1
  Yarn: 1.22.22
  pnpm: N/A
Relevant Packages:
  next: 14.2.5 // Latest available version is detected (14.2.5).
  eslint-config-next: 14.2.5
  react: 18.3.1
  react-dom: 18.3.1
  typescript: 5.5.4
Next.js Config:
  output: standalone

Which area(s) are affected? (Select all that apply)

Output (export/standalone), Performance

Which stage(s) are affected? (Select all that apply)

next start (local), Other (Deployed)

Additional context

I tested my reproduction code against different Next ~14 versions, canary releases and NodeJS versions. It looks like all the versions since async server components support are affected.
Also, I tested the same code with axios and cross-fetch, and it works correctly, so for me it looks like the issue is with Fetch caching.
I have the same issue in real-world app hosted in AKS, and it causes out of memory errors every few days - is there any way to make the garbage collector clean up memory allocations more aggressively when using native fetch?

I think I checked all the other similar issues like:
#64212
#54708
But I think that my example is much simpler so I decided to create new issue.

[hahmadzadeh]

We faced this problem in our platform, And we fixed it by downgrading node js to 20.15.1.

[khuezy]

Hey @kamdubiel I wonder if it's the console.log buffering the string that's causing the high memory usage. What if you removed the log?

[kamdubiel]

@khuezy I tested without console.log and it has no effect on memory usage.

@hahmadzadeh Thank you for that, indeed 20.15.1 works - I didn't test that one.

Unfortunately 22.6.0 that will become LTS in October is also affected.

[khuezy]

undici's fetch seems to be a nightmare. Have you notified the node team of the leaks happening to that library? There's this: https://undici.nodejs.org/#/?id=garbage-collection but your repo is consuming the body... so something else in node is leaking 🤷

[samcx]

@kamdubiel Thank you for submitting an issue and thanks for taking a look everyone!

Since this seems to be a Node issue and not a Next.js issue, I will be converting this to a discussion for further discussion in case anyone else comes across this.