Fetch failed error

[shawncao]

Description of the bug:

It seems having a high chance to hit failure, based on the error message, I'm not sure if this is SDK's issue or the model's issue. If you have some pointer, that would be great.

Often seeing this - more than 20% of the time? Since it says fetch failed, I'm wondering if this is something to do with the SDK or some client configurations.

[2025-02-27 11:47:16]: Failed to handle run status: Error: [GoogleGenerativeAI Error]: Error fetching from https://generativelanguage.googleapis.com/v1beta/models/gemini-2.0-flash:generateContent: fetch failed
[2025-02-27 11:47:32]: Chatting with user: [email protected]
history []
[2025-02-27 11:47:42]: Failed to handle run status: Error: [GoogleGenerativeAI Error]: Error fetching from https://generativelanguage.googleapis.com/v1beta/models/gemini-2.0-flash:generateContent: fetch failed

Actual vs expected behavior:

The service should be reliability give answer. But it often fails, my observation gives me ~20% of the chance.

Any other information you'd like to share?

No response

[hkt74]

Hi @shawncao

What is your runtime environment? I did some initial investigation, this error seems come from the low level fetch API. It looks like not from SDK (I couldn't find fetch failed error message from the repo) or API side (no http status code)

It would be helpful if you could provide the stack trace so we could debug further, thanks

[H9660]

Hi @shawncao could you explain how can we replicate this behaviour for further debugging?

[shawncao]

Hi, I can not repro this, this just randomly happened. I don't know if it is just the stability of Gemini service itself. If it is not related to the client lib, I can understand, we can close. Just want to bring it up to see if this is common for all users.

[0xdevalias]

There's not currently much extra info for it (hopefully they will provide some more details in time), but it looks like a humanify user may have just reported a similar issue:

• fetch failed jehna/humanify#366

It looks like they were running on Windows, and the stack trace seems to mention node:internal/deps/undici/undici:13502:13 as the source of the TypeError: fetch failed

Edit: They're running Node.js v22.14.0

I haven't looked deeply into this, but based on the error log you provided there, it seems to be happening within the @google/generative-ai package. Some potentially related issues from that upstream package:

• https://github.com/google-gemini/generative-ai-js/issues?q=sort%3Aupdated-desc%20is%3Aissue%20is%3Aopen%20%22fetch%20failed%22
  • TypeError: fetch failed #29

    • I was facing the same issue.

      I switched to node version 20 from 18, it started to work.

      Originally posted by @duke79 in TypeError: fetch failed #29 (comment)

    • There are also a lot of posts in this issue that seem to talk about this potentially being related to Chinese internet connections being unable to talk to Google; unsure if that is relevant here, but if so, then these issues may also be relevant:
      • Add support for proxy for openai/gemini modes jehna/humanify#68
      • How to set IP proxy? Requests to openai easily time out jehna/humanify#265
      • [Feature need]Support custom api_base jehna/humanify#212
  • Fetch failed error #360

From a google of "node internal deps undici", a few more potentially related issues popped up"

• fetch may try to use a closed connection nodejs/undici#3492
• internal undici:fetch throws error where node-fetch and curl succeed nodejs/node#45497

Originally posted by @0xdevalias in jehna/humanify#366 (comment)

[shawncao]

It's pretty random, feels like it's service reliability issue, once in a while, but I'm not sure. by the way, I'm running node v22.14.0

[hkt74]

Thanks @0xdevalias for sharing similar issues from other repos, particularly the stack trace shared in jehna/humanify#366 is very helpful. I also guess this might related to network connectivity issues, potentially due to firewall/proxy. (The issue happens randomly and hard to reproduce). If it is a Gemini api side issue, there should be a http status in the error message. I do found the same error message from the undici repo, which indicate it is a network error, without stack trace, it is hard to debug further.

@shawncao what do you think to close the issue for now and reopen when you have a detailed stack trace? thanks