kubernetes-sigs / lws / 41421845670
88%

DEFAULT BRANCH: main
Ran
Jobs 1
Files 41
Run time 1min
Badge
Embed ▾
README BADGES
x

If you need to use a raster PNG badge, change the '.svg' to '.png' in the link

Markdown

Textile

RDoc

HTML

Rst
coverage: 87.395% (-1.1%) from 88.501%
41421845670

push

github

web-flow 
feat: implement automatic ComputeDomain lifecycle management for MNNVL (KEP-1259) (#1263)

* feat: implement automatic ComputeDomain lifecycle management (KEP-1259)

Add opt-in NVIDIA DRA ComputeDomain (resource.nvidia.com/v1beta1) lifecycle
management and pod ResourceClaim injection for MNNVL workloads behind the
alpha AutoComputeDomain feature gate:

- Option 1 (per LeaderWorkerSet replica): leaderworkerset.sigs.k8s.io/auto-compute-domain=enabled
  allocates an isolated ComputeDomain per active replica across Ordinal and
  Hash groupIdentity modes, reconciled while the leader holds its scheduling
  gate and injected into GPU containers via PodWebhook.
- Option 2 (per DisaggregatedSet slice & revision): disaggregatedset.x-k8s.io/auto-compute-domain=enabled
  allocates one ComputeDomain per active (slice, revision) before child LWS
  creation, injecting DRA claims into enrolled roles while supporting role
  opt-out (disabled) and per-role replica-scope overrides.
- Container targeting: auto-detects containers and restartable sidecar
  initContainers requesting nvidia.com/gpu > 0, plus explicit override via
  leaderworkerset.sigs.k8s.io/compute-domain-containers for DRA GPU claims.
- Safe rollouts & teardown: upgrade-safe revision tracking (byte-identical
  revision hashes when unset/disabled), O(1) indexed Pod reference checks,
  and protection finalizer release once zero pods reference a domain.
- Conditional Helm RBAC via computeDomainManagement.enabled (default false).

Refs #1259

Signed-off-by: ahg-g <ahg@google.com>

* fix: harden ComputeDomain revision cache, gate toggles, and template annotations

- Include the effective auto-compute-domain value in the revision
  equality cache key. The annotation is part of the revision patch but
  does not bump the LWS generation, so a cached match could hide a
  toggle and leave leaders referencing a ComputeDomain that is never
  created.
- Stamp compute-domain-prefix and compute-domain-name on the leader ... (continued)

706 of 955 new or added lines in 15 files covered. (73.93%)

10275 of 11757 relevant lines covered (87.39%)

33008.97 hits per line

Uncovered Changes

Lines Coverage ∆ File
93
71.65
pkg/computedomain/reconciler.go
45
76.06
pkg/computedomain/injection.go
26
81.73
-3.74% pkg/controllers/disaggregatedset/disaggregatedset_controller.go
25
91.45
-3.8% pkg/webhooks/disaggregatedset/disaggregatedset_webhook.go
18
85.57
-1.13% pkg/controllers/leaderworkerset_controller.go
15
29.74
-1.3% cmd/main.go
7
82.82
-0.36% pkg/controllers/pod_controller.go
5
88.05
1.73% pkg/webhooks/leaderworkerset_webhook.go
4
94.93
-0.41% pkg/controllers/disaggregatedset/lws_manager.go
4
89.43
-0.4% pkg/utils/revision/revision_utils.go
3
83.14
-0.67% pkg/controllers/leaderworkerset_hash.go
2
97.31
-0.74% pkg/utils/disaggregatedset/utils.go
2
90.09
-0.7% pkg/webhooks/pod_webhook.go
Jobs
ID Job ID Ran Files Coverage
1 41421845670.1 41
87.39
GitHub Action Run
Source Files on build 41421845670