Enhances runner controller and runner pod controller to have consistent timeouts for runner unregistration and runner pod deletion,
so that we are very much unlikely to terminate pods that are running any jobs.
There is a race condition between ARC and GitHub service about deleting runner pod.
- The ARC use REST API to find a particular runner in a pod that is not running any jobs, so it decides to delete the pod.
- A job is queued on the GitHub service side, and it sends the job to this idle runner right before ARC deletes the pod.
- The ARC delete the runner pod which cause the in-progress job to end up canceled.
To avoid this race condition, I am calling `r.unregisterRunner()` before deleting the pod.
- `r.unregisterRunner()` will return 204 to indicate the runner is deleted from the GitHub service, we should be safe to delete the pod.
- `r.unregisterRunner` will return 400 to indicate the runner is still running a job, so we will leave this runner pod as it is.
TODO: I need to do some E2E tests to force the race condition to happen.
Ref #911
Apparently, we've been missed taking an updated registration token into account when generating the pod template hash which is used to detect if the runner pod needs to be recreated.
This shouldn't have been the end of the world since the runner pod is recreated on the next reconciliation loop anyway, but this change will make the pod recreation happen one reconciliation loop earlier so that you're less likely to get runner pods with outdated refresh tokens.
Ref https://github.com/actions-runner-controller/actions-runner-controller/pull/1085#issuecomment-1027433365
This allows providing a different `work` Volume.
This should be a cloud agnostic way of allowing the operator to use (for example) NVME backed storage.
This is a working example where the workDir will use the provided volume, additionally here docker is placed on the same NVME.
```
apiVersion: actions.summerwind.dev/v1alpha1
kind: RunnerDeployment
metadata:
name: runner-2
spec:
template:
spec:
dockerdContainerResources: {}
env:
- name: POD_NAME
valueFrom:
fieldRef:
fieldPath: metadata.name
# this is to mount the docker in docker onto NVME disk
dockerVolumeMounts:
- mountPath: /var/lib/docker
name: scratch
subPathExpr: $(POD_NAME)-docker
- mountPath: /runner/_work
name: work
subPathExpr: $(POD_NAME)-work
volumeMounts:
- mountPath: /runner/_work
name: work
subPathExpr: $(POD_NAME)-work
dockerEnv:
- name: POD_NAME
valueFrom:
fieldRef:
fieldPath: metadata.name
volumes:
- hostPath:
path: /mnt/disks/ssd0
name: scratch
- hostPath:
path: /mnt/disks/ssd0
name: work
nodeSelector:
cloud.google.com/gke-nodepool: runner-16-with-nvme
ephemeral: false
image: ""
imagePullPolicy: Always
labels:
- runner-2
- self-hosted
organization: yourorganization
```
* fix(deps): update module sigs.k8s.io/controller-runtime to v0.11.0
* Fix dependencies and bump Go to 1.17 so that it builds after controller-runtime 0.11.0 upgrade
* Regenerate manifests with the latest K8s dependencies
Co-authored-by: Renovate Bot <bot@renovateapp.com>
Co-authored-by: Yusuke Kuoka <ykuoka@gmail.com>
* Adding a default docker registry mirror
This change allows the controller to start with a specified default
docker registry mirror and avoid having to specify it in all the runner*
objects.
The change is backward compatible, if a runner has a docker registry
mirror specified, it will supersede the default one.
* Add POC of GitHub Webhook Delivery Forwarder
* multi-forwarder and ctrl-c existing and fix for non-woring http post
* Rename source files
* Extract signal handling into a dedicated source file
* Faster ctrl-c handling
* Enable automatic creation of repo hook on startup
* Add support for forwarding org hook deliveries
* Set hook secret on hook creation via envvar (HOOK_SECRET)
* Fix org hook support
* Fix HOOK_SECRET for consistency
* Refactor to prepare for custom log position provider
* Refactor to extract inmemory log position provider
* Add configmap-based log position provider
* Rename githubwebhookdeliveryforwarder to hookdeliveryforwarder
* Refactor to rename LogPositionProvider to Checkpointer and extract ConfigMap checkpointer into a dedicated pkg
* Refactor to extract logger initialization
* Add hookdeliveryforwarder README and bump go-github to unreleased ver
This option expose internally some `KUBERNETES_*` environment variables
that doesn't allow the runner to use KinD (Kubernetes in Docker) since it will
try to connect to the Kubernetes cluster where the runner it's running.
This option it's set by default to `true` in any Kubernetes deployment.
Signed-off-by: Jonathan Gonzalez V <jonathan.gonzalez@enterprisedb.com>
* feat: RunnerSet backed by StatefulSet
Unlike a runner deployment, a runner set can manage a set of stateful runners by combining a statefulset and an admission webhook that mutates statefulset-managed pods with required envvars and registration tokens.
Resolves#613
Ref #612
* Upgrade controller-runtime to 0.9.0
* Bump Go to 1.16.x following controller-runtime 0.9.0
* Upgrade kubebuilder to 2.3.2 for updated etcd and apiserver following local setup
* Fix startup failure due to missing LeaderElectionID
* Fix the issue that any pods become unable to start once actions-runner-controller got failed after the mutating webhook has been registered
* Allow force-updating statefulset
* Fix runner container missing work and certs-client volume mounts and DOCKER_HOST and DOCKER_TLS_VERIFY envvars when dockerdWithinRunner=false
* Fix runnerset-controller not applying statefulset.spec.template.spec changes when there were no changes in runnerset spec
* Enable running acceptance tests against arbitrary kind cluster
* RunnerSet supports non-ephemeral runners only today
* fix: docker-build from root Makefile on intel mac
* fix: arch check fixes for mac and ARM
* ci: aligning test data format and patching checks
* fix: removing namespace in test data
* chore: adding more ignores
* chore: removing leading space in shebang
* Re-add metrics to org hra testdata
* Bump cert-manager to v1.1.1 and fix deploy.sh
Co-authored-by: toast-gear <15716903+toast-gear@users.noreply.github.com>
Co-authored-by: Callum James Tait <callum.tait@photobox.com>
Sets the privileged flag to false if SELinuxOptions are present/defined. This is needed because containerd treats SELinux and Privileged controls as mutually exclusive. Also see https://github.com/containerd/cri/blob/aa2d5a97c/pkg/server/container_create.go#L164.
This allows users who use SELinux for managing privileged processes to use GH Actions - otherwise, based on the SELinux policy, the Docker in Docker container might not be privileged enough.
Signed-off-by: Jonah Back <jonah@jonahback.com>
Co-authored-by: Yusuke Kuoka <ykuoka@gmail.com>
This allows using the `runtimeClassName` directive in the runner's spec.
One of the use-cases for this is Kata Containers, which use `runtimeClassName` in a pod spec as an indicator that the pod should run inside a Kata container. This allows us a greater degree of pod isolation.
- Adds `ephemeral` option to `runner.spec`
```
....
template:
spec:
ephemeral: false
repository: mumoshu/actions-runner-controller-ci
....
```
- `ephemeral` defaults to `true`
- `entrypoint.sh` in runner/Dockerfile modified to read `RUNNER_EPHEMERAL` flag
- Runner images are backward-compatible. `--once` is omitted only when the new envvar `RUNNER_EPHEMERAL` is explicitly set to `false`.
Resolves#457
This is an attempt to support scaling from/to zero.
The basic idea is that we create a one-off "registration-only" runner pod on RunnerReplicaSet being scaled to zero, so that there is one "offline" runner, which enables GitHub Actions to queue jobs instead of discarding those.
GitHub Actions seems to immediately throw away the new job when there are no runners at all. Generally, having runners of any status, `busy`, `idle`, or `offline` would prevent GitHub actions from failing jobs. But retaining `busy` or `idle` runners means that we need to keep runner pods running, which conflicts with our desired to scale to/from zero, hence we retain `offline` runners.
In this change, I enhanced the runnerreplicaset controller to create a registration-only runner on very beginning of its reconciliation logic, only when a runnerreplicaset is scaled to zero. The runner controller creates the registration-only runner pod, waits for it to become "offline", and then removes the runner pod. The runner on GitHub stays `offline`, until the runner resource on K8s is deleted. As we remove the registration-only runner pod as soon as it registers, this doesn't block cluster-autoscaler.
Related to #447
Changes:
- Switched to use `jq` in startup.sh
- Enable docker registry mirror configuration which is useful when e.g. avoiding the Docker Hub rate-limiting
Check #478 for how this feature is tested and supposed to be used.
Enable the user to set a limit size on the volume of the runner to avoid some runner pod affecting other resources of the same cluster
Co-authored-by: Yusuke Kuoka <ykuoka@gmail.com>
This makes logging more concise by changing logger names to something like `controllers.Runner` to `actions-runner-controller.runner` after the standard `controller-rutime.controller` and reducing redundant logs by removing unnecessary requeues. I have also tweaked log messages so that their style is more consistent, which will also help readability. Also, runnerreplicaset-controller lacked useful logs so I have enhanced it.
Since #392, the runner controller could have taken unexpectedly long time until it finally notices that the runner has been registered to GitHub. This patch fixes the issue, so that the controller will notice the successful registration in approximately 1 minute(hard-coded).
More concretely, let's say you had configured a long sync-period of like 10m, the runner controller could have taken approx 10m to notice the successful registration. The original expectation was 1m, because it was intended to recheck every 1m as implemented in #392. It wasn't working as such due to my misunderstanding in how requeueing work.
We occasionally encountered those errors while the underlying RunnerReplicaSet is being recreated/replaced on RunnerDeployment.Spec.Template update. It turned out to be due to that the RunnerDeployment controller was waiting for the runner pod becomes `Running`, intead of the new replacement runner to have registered to GitHub. This fixes that, by trying to Runner.Status.Phase to `Running` only after the runner in the runner pod appears to be registered.
A side-effect of this change is that runner controller would call more "ListRunners" GitHub Actions API. I've reviewed and improved the runner controller code and Runner CRD to make make the number of calls minimum. In most cases, ListRunners should be called only twice for each runner creation.
* if a runner pod starts up with an invalid token, it will go in an
infinite retry loop, appearing as RUNNING from the outside
* normally, this error situation is detected because no corresponding
runner objects exists in GitHub and the pod will get removed after
registration timeout
* if the GitHub runner object already existed before - e.g. because a
finalizer was not properly run as part of a partial Kubernetes crash,
the runner will always stay in a running mode, even updating the
registration token will not kill the problematic pod
* introducing RunnerOffline exception that can be handled in runner
controller and replicaset controller
* as runners are offline when a pod is completed and marked for restart,
only do additional restart checks if no restart was already decided,
making code a bit cleaner and saving GitHub API calls after each job
completion
* the reconciliation loop is often much faster than the runner startup,
so changing runner not found related messages to debug and also add the
possibility that the runner just needs more time
* errors.Is compares all members of a struct to return true which never
happened
* switched to type check instead of exact value check
* notRegistered was using double negation in if statement which lead to
unregistering runners after the registration timeout
* if a k8s node becomes unresponsive, the kube controller will soft
delete all pods after the eviction time (default 5 mins)
* as long as the node stays unresponsive, the pod will never leave the
last status and hence the runner controller will assume that everything
is fine with the pod and will not try to create new pods
* this can result in a situation where a horizontal autoscaler thinks
that none of its runners are currently busy and will not schedule any
further runners / pods, resulting in a broken runner deployment until the
runnerreplicaset is deleted or the node comes back online
* introducing a pod deletion timeout (1 minute) after which the runner
controller will try to reboot the runner and create a pod on a working
node
* use forceful deletion and requeue for pods that have been stuck for
more than one minute in terminating state
* gracefully handling race conditions if pod gets finally forcefully deleted within
