|
| 1 | +<!-- |
| 2 | +Copyright The Shipwright Contributors |
| 3 | +
|
| 4 | +SPDX-License-Identifier: Apache-2.0 |
| 5 | +--> |
| 6 | + |
| 7 | +--- |
| 8 | +title: build-runtime-class |
| 9 | +authors: |
| 10 | + - "@adambkaplan" |
| 11 | +reviewers: |
| 12 | + - "@dorzel" |
| 13 | + - "@HeavyWombat" |
| 14 | +approvers: |
| 15 | + - "@qu1queee" |
| 16 | + - "@SaschaSchwarze0" |
| 17 | +creation-date: 2025-03-25 |
| 18 | +last-updated: 2025-03-25 |
| 19 | +status: provisional |
| 20 | +see-also: |
| 21 | + - "/ships/0039-build-scheduler-opts.md" |
| 22 | +replaces: [] |
| 23 | +superseded-by: [] |
| 24 | +--- |
| 25 | + |
| 26 | +# Build Runtime Class |
| 27 | + |
| 28 | +## Release Signoff Checklist |
| 29 | + |
| 30 | +- [ ] Enhancement is `implementable` |
| 31 | +- [ ] Design details are appropriately documented from clear requirements |
| 32 | +- [ ] Test plan is defined |
| 33 | +- [ ] Graduation criteria for dev preview, tech preview, GA |
| 34 | +- [ ] User-facing documentation is created in [docs](/docs/) |
| 35 | + |
| 36 | +## Open Questions [optional] |
| 37 | + |
| 38 | +- Should user namespaces also be in scope for this feature? |
| 39 | + _No, this should be considered separately. Feature is beta as of k8s v1.30_. |
| 40 | + |
| 41 | +## Summary |
| 42 | + |
| 43 | +Extend the `Build` and `BuildRun` APIs to let build pods select their [RuntimeClass](https://kubernetes.io/docs/concepts/containers/runtime-class/) |
| 44 | +for execution. |
| 45 | + |
| 46 | +## Motivation |
| 47 | + |
| 48 | +The `RuntimeClass` API lets Kubernetes clusters support multiple container runtime environments. |
| 49 | +This is most often encountered when using [Kata containers](https://katacontainers.io/), which adds |
| 50 | +hardware virtualization to the existing mechanisms for isolating containers. |
| 51 | + |
| 52 | +### Goals |
| 53 | + |
| 54 | +- Allow builds to run with alternative container runtimes, such as Kata containers. |
| 55 | + |
| 56 | +### Non-Goals |
| 57 | + |
| 58 | +- Automatic mechanisms for selecting the container runtime. |
| 59 | +- Network isolated builds. |
| 60 | +- Isolation with Kubernetes [user namespaces](https://kubernetes.io/docs/concepts/workloads/pods/user-namespaces/). |
| 61 | + |
| 62 | +## Proposal |
| 63 | + |
| 64 | +### User Stories |
| 65 | + |
| 66 | +#### Kata Containers |
| 67 | + |
| 68 | +As a developer building containers with Shipwright I want to specify that the build pod containers |
| 69 | +use Kata containers as their runtime so that I can isolate my builds with hardware virtualization. |
| 70 | + |
| 71 | +### Implementation Notes |
| 72 | + |
| 73 | +#### API Update |
| 74 | + |
| 75 | +The `BuildSpec` API for Build and BuildRun will be updated to add the `runtimeClass` field: |
| 76 | + |
| 77 | +```yaml |
| 78 | +spec: |
| 79 | + ... |
| 80 | + runtimeClassName: kata # string |
| 81 | +``` |
| 82 | +
|
| 83 | +This field will correspond to the [runtimeClassName](https://kubernetes.io/docs/reference/kubernetes-api/workload-resources/pod-v1/#scheduling) |
| 84 | +field on Kubernetes pods. |
| 85 | +
|
| 86 | +#### Precedence Ordering and Value Merging |
| 87 | +
|
| 88 | +Values in `BuildRun` will override those in the referenced `Build` object (if present). This allows |
| 89 | +the `BuildRun` object to "inherit" values from a parent `Build` object. |
| 90 | + |
| 91 | +#### Impact on Tekton TaskRun |
| 92 | + |
| 93 | +Tekton supports tuning the pod of the `TaskRun` using the |
| 94 | +[podTemplate](https://tekton.dev/docs/pipelines/taskruns/#specifying-a-pod-template) field. When |
| 95 | +Shipwright creates the `TaskRun` for a build, the respective `runtimeClassName` can be passed |
| 96 | +through. |
| 97 | + |
| 98 | +#### Command Line Enhancements |
| 99 | + |
| 100 | +The `shp` CLI _may_ be enhanced to add flags that set the `runtimeClassName` for a `Build` and/or |
| 101 | +`BuildRun`. For example, `shp build run` can have the following new options: |
| 102 | + |
| 103 | +- `--runtime-class=<value>`: this would set the respective `spec.runtimeClassName` value on the |
| 104 | + generated `BuildRun`. |
| 105 | + |
| 106 | +### Test Plan |
| 107 | + |
| 108 | +- Unit testing can verify that the generated `TaskRun` object for a build contains the desired pod |
| 109 | + template fields. |
| 110 | +- End to end testing may prove challenging, as KinD is not designed to support multiple container |
| 111 | + runtimes. Implementations like Kata Containers are designed for real clusters on cloud provider |
| 112 | + infrastructure. |
| 113 | + |
| 114 | + |
| 115 | +### Release Criteria |
| 116 | + |
| 117 | +**Note:** *Section not required until targeted at a release.* |
| 118 | + |
| 119 | +#### Removing a deprecated feature [if necessary] |
| 120 | + |
| 121 | +Not applicable. |
| 122 | + |
| 123 | +#### Upgrade Strategy [if necessary] |
| 124 | + |
| 125 | +The `runtimeClassName` field will be optional and default to Golang empty values. |
| 126 | +On upgrade, these values will remain empty on existing `Build`/`BuildRun` objects. |
| 127 | + |
| 128 | +### Risks and Mitigations |
| 129 | + |
| 130 | +#### Pod Scheduling Collision |
| 131 | + |
| 132 | +The `RuntimeClass` object (set and maintained by cluster admins/platform teams) also has options |
| 133 | +to set defaults for [pod scheduling](https://kubernetes.io/docs/concepts/containers/runtime-class/#scheduling), |
| 134 | +such as node selectors and tolerations. As of Kubernetes 1.32, this is a "beta" feature enabled |
| 135 | +on most Kubernetes distributions by default. These scheduler options are merged on pod admission; |
| 136 | +there is a risk that pods are rejected and builds fail due to colliding pod scheduler values. |
| 137 | + |
| 138 | +In theory the controller for `BuildRuns` can anticipate this situation and fail the build quickly. |
| 139 | +However, other Kubernetes controllers do not appear to do this, and thus it may not be worthwhile |
| 140 | +to add this extra logic. Checking the `RuntimeClass` for every build also adds overhead to each |
| 141 | +reconcile loop, either by making kubernetes apiserver calls directly or adding an informer-based |
| 142 | +cache for `RuntimeClass` that provides little value. |
| 143 | + |
| 144 | +## Drawbacks |
| 145 | + |
| 146 | +This continues our leaking of Kubernetes pod APIs to Shipwright - see |
| 147 | +[SHIP-0039](https://github.com/shipwright-io/community/blob/main/ships/0039-build-scheduler-opts.md#drawbacks). |
| 148 | +Similar drawbacks also exist with respect to cluster admin control over pod scheduling. |
| 149 | + |
| 150 | +## Alternatives |
| 151 | + |
| 152 | +### Kubernetes User Namespaces |
| 153 | + |
| 154 | +An initial draft of this proposal included Kubernetes |
| 155 | +[user namespaces](https://kubernetes.io/docs/concepts/workloads/pods/user-namespaces/) as part of |
| 156 | +the scope. This was removed because user namespaces are a beta feature in Kubernetes (as of v1.32), |
| 157 | +and did not reach the beta milestone until v1.30. Even with achievement of the beta milestone, the |
| 158 | +feature is still disabled by default in Kubernetes v1.32, making it challenging to test. |
| 159 | + |
| 160 | +## Infrastructure Needed [optional] |
| 161 | + |
| 162 | +No additional infrastructure anticipated. |
| 163 | + |
| 164 | + |
| 165 | +## Implementation History |
| 166 | + |
| 167 | +- 2025-03-25: Created as `provisional` |
0 commit comments