Remaining fixes needed to CI on AKS #18
Description
With PR #16, automatic CI on AKS is enabled, with PerformanceTestInterruptible running between two pods of a Kubernetes cluster. However, there are two limitations with how it is currently set up:
- The test is currently running the ambrosia/ambrosia-perftest public docker image, which is wrong, it should test the version of PerformanceTestInterruptible that is contained in the given commit of the AMBROSIA repo. It's testing the latest version of the AKS scripts, but not the latest version of PTI itself.
- The test currently cannot fully reprovision the Azure resources. It runs fine when the resources are already there (idempotent) but cannot create each of them. That required pre-population by a local run of the script.
Both of these problems are authentication related. The ADO pipeline needs to authenticate with Dockerhub if it's going to execute a docker push. I allocated the ambrosia/ambrosia-ci-test repository for just this purpose, but I ran out of time before figuring out the ADO->Dockerhub auth.
The second problem is related to the permissions and powers of the CI user vis a vis the Azure CLI calls used in the AKS-scripts directory. The AKS-scripts/run-end-to-end* will provision everything from scratch for a local user. But if the resource group is wiped away, then it will fail to provision everything inside a CI job.
(Overcoming that is even more challenging if one wants to push to a private Azure Container Registry instead of public dockerhub.)