Your own cluster
Every run starts by choosing one of your organisation’s clusters, in this order. The first that answers wins.
A production cluster, or one whose purpose nobody confirmed, is never chosen by the last rule, even when it is the organisation’s only cluster. With no cluster at all, a push, pull request, tag or schedule starts no run and gets a check titled Choose a CI cluster for this organisation. A run started by hand still starts: its
build steps can go to Ankra’s build cluster, and its other steps conclude with that reason instead of running.
A repository can also send its runs to one cluster of its own, whatever the organisation’s settings say. Its next runs go there; runs already queued keep theirs:
ankra-ci namespaces, on its CI node group if it has one. The cluster’s agent needs pipeline workers: each worker runs one step at a time. Your source is checked out, built and tested there.
What it costs: the nodes your cluster already runs on. A step reserves the CPU and memory it asks for, so a busy hour needs room on those nodes, or a node group that can grow.
Ankra’s build cluster
Abuild step builds a container image. Three things stop a run’s own cluster from building one: the run has no cluster at all (a run started by hand, above), the cluster’s agent does not run pipeline steps, or the nodes confine the rootless image builder (AppArmor or seccomp, common on Ubuntu and k3s nodes), which a step reports as build_runtime_confined. When that happens, Ankra can build the image on its own build cluster instead. A build for a pull request from a fork never goes there.
- Only the builder is confined. Only
buildmoves. Every other step runs on your cluster as usual. - No cluster, or an agent that runs no pipeline steps. The run’s
scansteps run on Ankra’s platform scanner, against the commit and the image, so the scanner receives your source too. Steps that need your cluster, such as your own tests, are refused.gateandpublishare settled by Ankra as usual.
- The organisation’s build fallback is on. It is on until an owner or admin turns it off: Build on Ankra when your cluster cannot, under Build rules in Build & delivery → Builds and clusters. In the CLI it is
--build-fallback platform_builders. - Ankra has enabled platform builds for the organisation. This is Ankra’s grant, not a setting you can change. It is on for every organisation unless Ankra turned it off for yours.
ankra org ci-settings getshows it.
A CI playground
A run whose largest step no node of your CI cluster can hold would wait until its time ran out. A cluster where a step recently could not be scheduled, or was killed at its memory limit, in the last seven days, counts as too small too. For those runs, Ankra moves the whole run to a CI playground: an Ankra Cloud cluster it builds for the run, with pipeline workers. While the run waits for it, its queue reason iswaiting_for_ci_playground.
- The cheapest arrangement wins. When the run’s largest step fits Ankra Cloud’s trial plan, the organisation’s own free trial playground takes it: one already running, or a new one for an account without a payment method that still has a trial. Otherwise, when the organisation’s Ankra Cloud account can pay, it gets a billed playground at the smallest Ankra Cloud plan whose node fits the run’s largest step. An account linked to an Ankra organisation usually pays only from the credit the organisation tops up, so a card alone may not be enough. When Ankra Cloud refuses the playground, for example an account with no payment method and no trial left, the run stays on your cluster.
- One playground per organisation. Every run that needs one shares it and queues for room on it like on any CI cluster. It grows to a larger plan only for a single step it cannot hold, and only once no step is running on it, because a resize restarts its server.
- A billed playground is removed when it goes idle. Once no run has used it for the organisation’s idle time, 5 minutes unless you change it, Ankra tears it down. A trial playground is left to its own expiry.
resources, or add a node group with bigger nodes.
Keep everything in your own cluster
Two organisation settings keep every pipeline step on your own clusters. An organisation owner or admin changes them.ankra org ci-settings set --playground-idle-minutes 30, or {"ci_playground_idle_minutes": 30} in the same API call, changes how long an idle billed playground is kept, from 0 to 1440 minutes. A value outside that range is refused; the API answers “CI playground idle time must be between 0 and 1440 minutes.”
ankra org ci-settings get prints Build fallback and Platform builds enabled, and from CLI 0.78.0 CI playground and Playground idle time as well. GET https://platform.ankra.app/api/v1/org/ci-settings returns ci_build_fallback, ci_auto_playground and ci_playground_idle_minutes.
Builds outside a pipeline run
The two settings cover pipeline steps. Ankra can also build an application’s image on its build cluster outside a pipeline run, for example the first build when you create the application, or a pull request preview whose image is missing. Ankra starts those builds by itself only while the build fallback is on and the organisation has Ankra’s platform builds grant, whichankra org ci-settings get shows as Platform builds enabled. With the build fallback off it starts none of them.
If your source may never be built outside your own infrastructure, turn both settings off before you create an application. A build someone starts on Ankra’s build cluster by name, with ankra application build start or ankra application ship with --ankra-build, still runs there.
What Ankra keeps about a run
Wherever a step runs, on your cluster, Ankra’s build cluster or a CI playground, Ankra keeps the same things about it:
Ankra counts what it holds in the Ankra-managed default vault; nothing in a vault of your own is counted. Change the retention on Build & delivery → Builds and clusters, under How long to keep pipeline results.
Related
- Move to Ankra Pipelines: choose the cluster and its workers once for the organisation.
- Troubleshooting: a step that waits, or a build that fails before the Dockerfile is read.
- Cluster autoscaling: let the CI node group grow instead of queueing steps.