Hi @saudet ,

I checked your workflow yml file, you are using container nvidia/cuda:10.1-cudnn7-devel-centos7 to run your job. 

Based on my test , the decreased disk space is used by this image. Please see my job yml and workflow logs :

job1:
    runs-on: ubuntu-latest
    steps:
      - name: df
        run: |
          df -h
          docker system df
          docker run nvidia/cuda:10.1-cudnn7-devel-centos7
          df -h
          docker system df

before.png733×574 13.9 KB

after.png757×540 14.9 KB

The size of image is 3.5GB . 

Yes, I know and I did mention that the container takes about 4 GB of space. The point is, 13 GB of space is less than the 14 GB that is supposed to be guaranteed:

Each virtual machine has the same hardware resources available.

• 2-core CPU
• 7 GB of RAM memory
• 14 GB of SSD disk space

https://help.github.com/en/actions/reference/virtual-environments-for-github-hosted-runners

If you only guarantee 13 GB, then the documentation should be updated. Do you agree?

As of at least yesterday, we’re now down to about 9 GB:

https://github.com/tensorflow/java/runs/599744540 (Apr 20, 2020)

overlay 84G 80G 3.7G 96% /

I doubt very much we’re supposed to have so little disk space to work with.

I had seen the same thing on our builds that started to fail. The following values are after docker system prune:

/dev/sda1    84G    74G    9.9G    89%    /

And today:

/dev/sda1    84G    70G    15G    84%    /

So they fixed whatever they temporarely broke (I hope). It would have been nice to acknowledge the problem, so that we now that a fix was underway.

You could see the logs in my above screenshot , the /dev/sdb1 is 14 G, used 41M, then the Avail show 13G . It is not an accurate data.