We’re using Google Batch to run some of our predictions, and our pipeline failed overnight because the head of the official image family doesn’t advertised any guest OS features (while all other previous images do) which means the submitted batch jobs just fail with:
VM in Managed Instance Group meets error: Batch Error: code - CODE_GCE_INVALID_FIELD_VALUE, description - error count is 6, latest message example: Instance ‘image-name-redacted’ creation failed: Invalid value for field ‘resource.machineType’: ‘projects/project-number-redacted/zones/europe-west2-a/machineTypes/c4-highcpu-96’. NetworkInterface NicType can only be set to GVNIC on instances with GVNIC GuestOsFeature.
The command to get the images and their guest OS features (in TYPE):
gcloud compute images list --project batch-custom-image --filter=“family=batch-cos-stable-official” --format=“table(name,creationTimestamp,guestOsFeatures[].type.list())” --show-deprecated --sort-by=~creationTimestamp
It looks like the issue is tied specifically to the newer COS image family and the GVNIC GuestOsFeature requirement rather than Batch itself. Since the older images are still working, I’d compare the GuestOsFeatures and machine type/network interface requirements between the last working image and the new 20260825 image.
I’d also check whether the image metadata changed between those releases. If the latest image is advertising a network interface requirement that the selected machine type doesn’t support, that would explain why the job fails during instance creation rather than later in the boot process.
Since you already have the image list and GuestOsFeatures output, comparing the last known-good image against the new one seems like the quickest way to narrow this down.
In the OP’s message, the metadata for the image is provided, showing the latest image is missing the `GVNIC GuestOSFeature`: something is clearly broken. I experienced an issue with the batch COS image earlier this year, which was confirmed as a Google problem: Cloud Batch job passing unexpected arguments to runnables - #4 by Lobna_Mazhar .
Can someone from Google look into this? Maybe @Lobna_Mazhar? And advise whether or not this product can be reliably used in production? All signs currently point to “no”.
I completely agree with @output4616. We are having this issue in production every now and then (and today like OP). Fortunately we can fix easily but that’s really a pain.
> All signs currently point to “no”.
+1, + when you compare Batch UI vs Cloud Run’s you can definitely see that there’s one that gets all the investments (the one that is marketed) and one that does not
There was an issue with our latest images dated 24th August 2026, which caused them to drop the guest OS features. These images have now been deprecated and the previous versions, dated 6th August, have been made the default ones.
We also know what the root cause was, so the next version of images will have the guest OS features populated.