The actual implementation fails when Quay is processing layers bigger than 4GiB because the upload on GCS is not multi-part. Therefore, Quay collects the whole layer in the memory and uploads it entirely on GCS.

Problems are two:

• Collecting the whole layer it's not possible when Quay it's deployed on Openshift because it has a memory limit of 8Gib and gunicorn process gets killed for OOM. The workaround for this is to scale down the Quay operator and to modify manually the Quay-App deployment.
• The whole layer on GCS needs to be pushed within 60 seconds. The workaround for this is described here https://issues.redhat.com/browse/PROJQUAY-6819

The two workarounds exist or at least they will be sooner or later available for every customer that uses Quay on GCP. However, the memory issue can be problematic when working with more than one "large" image simultaneously. Therefore, a reorg of the Quay push on GCS should be evaluated.

Useful links:

• https://cloud.google.com/storage/docs/multipart-uploads
• Suggested method by Google: https://cloud.google.com/storage/docs/resumable-uploads