Deploy fasttext model downloading error
We can build the app correctly, but when we deploy and want to download fasttext pre-trained model, the following error occurs:
[2023-11-20 03:05:06 +0000] [1] [CRITICAL] WORKER TIMEOUT (pid:2404)
[2023-11-20 03:05:07 +0000] [2467] [INFO] Booting worker with pid: 2467
Solution:Jump to solution
have another python file that only downloads the model, like a download_model.py file, and then change your start command to something like this
python download_model.py && <your current start command>17 Replies
Project ID:
5a2cb014-5557-4588-976f-dab9ac88315a5a2cb014-5557-4588-976f-dab9ac88315a
Timeout
Lol but your worker timeouts before finishing download
Increase workers life
Hi! how can i do that?
Brody can help, you might need custom nginx config no idea how it works on railway
dont give up that easy
Lol
well the real question is, why is the worker downloading the model?
sure we could increase the timeout, but that would only be a band aid fix
mmm how would you approach the problem? We need to download the model each time we deploy the app, so we have the model stored to do predictions.
Solution
have another python file that only downloads the model, like a download_model.py file, and then change your start command to something like this
python download_model.py && <your current start command>Whats the difference between doing it at the start or when the app is running? Is the worker not the same? or im missing something?
a worker is a gunicorn web worker, it handles the incoming http requests, they should not be downloading a model
Using worker is fine but you need to be more sophisticated around streaming large files
Stream in chunks
plenty of reasons why not, but heres one reason, maybe you want to increase the worker count in the future to handle more traffic, you add 4 workers, then you will be downloading the model 4 times when you only need to download it once
pandas, the app is downloading the model, not the client
That's why you have to verify chunks and check which is downloaded
Totally viable to do just a lot harder than adding simple download script before app initiates