Cloud Run jobs cannot pull arbitrary Docker Hub images: "Image mirror.gcr.io/... not found"

Goal: get past a Compute Engine CPU quota (CPUS_ALL_REGIONS 12 on a new project) by running batch CFD as Cloud Run jobs, which have separate capacity.

Symptom: gcloud run jobs create ... --image=docker.io/openfoam/openfoam8-paraview56:latest failed with "Image mirror.gcr.io/openfoam/openfoam8-paraview56:latest not found".

Cause: Cloud Run rewrites docker.io references to mirror.gcr.io, Google's pull-through cache, which only holds commonly used images. Less popular Docker Hub images must be copied into Artifact Registry first.

What that takes: enable artifactregistry.googleapis.com and cloudbuild.googleapis.com, create a docker repository, then gcloud builds submit DIR --tag REGION-docker.pkg.dev/PROJECT/REPO/IMAGE:TAG with a Dockerfile whose FROM line names the Docker Hub image (Cloud Build can pull from Docker Hub). On a fresh personal project this failed with PERMISSION_DENIED for the user account, after the build had already created a PROJECT_cloudbuild staging bucket. Clean that up if you abandon the attempt.

Other notes: - Cloud Run jobs can mount a Cloud Storage bucket (--add-volume=name=gcs,type=cloud-storage,bucket=B --add-volume-mount=volume=gcs,mount-path=/mnt/gcs). Their local filesystem is in memory, so large CFD output counts against the memory limit (up to 32 GiB with 8 vCPU). - Give the job a dedicated service account with only bucket-level roles/storage.objectAdmin, rather than the default compute account. - Cloud Batch and GKE use Compute Engine VMs, so they do not get around the CPUS_ALL_REGIONS quota. Cloud Run does, if you can get the image into Artifact Registry. - gcloud list flags split on commas. For a long bash -c script in --args, use the alternate delimiter syntax: --args="^@^-c@your script, with commas".