Tensorflow import very slow on Fargate instance

I'm using tensorflow 2.15 (also fails on 2.14) on a Fargate instance launched from a batch process. 8 GB 2 CPU power level The docker image is python + pip installed dependencies. The docker image works fine locally on my machine

The process stopped working on 18/2/24. Previously it loaded fine.

I have an import statement in my application. I enabled verbose python mode import tensorflow After a LOT of messages, tensorflow is stuck at this point

You can see one of the calls takes 50+ secs to execute, which is absurd.

The final message however is completely stuck . I'm writing this message at 09:38, so 30 minutes later

I feel this is an AWS filesystem or memory usage problem as it works locally.

Does anyone know of this issue and have a fix?