mirror of
https://github.com/hanzo-ml/trainer.git
synced 2026-09-24 00:31:34 +00:00
Sabari NarayanaandGitHub
844656a01d
* fix(webhook): allow atomic RuntimePatches update on unsuspend Change checkRuntimePatchesImmutability to check the *old* TrainJob's suspend state instead of the new one. This lets external controllers (notably Kueue, see kubernetes-sigs/kueue#8296) modify spec.runtimePatches and set spec.suspend=false in a single API request, eliminating the two-step update workaround and the race window between the calls. Mirrors the Kubernetes core Job validation pattern: https://github.com/kubernetes/kubernetes/blob/86b66f6f333a/pkg/registry/batch/job/strategy.go#L191 The downstream JobSet activity check (no active replicatedJobs) still runs, so updates while pods are running remain forbidden. Refs: #3043 Signed-off-by: Sabari Narayana <[email protected]> * test(jobset): cover atomic suspend->unsuspend RuntimePatches update Add a unit test that exercises the new atomic-update path: a single TrainJob update that both clears spec.suspend and modifies spec.runtimePatches must succeed when the previous state was suspended. Refs: #3043 Signed-off-by: Sabari Narayana <[email protected]> * test(integration): cover atomic unsuspend + RuntimePatches update Extend the trainjob webhook update table with an entry that creates a suspended TrainJob with RuntimePatches and then, in a single Update, both unsuspends it and modifies the runtimePatches NodeSelector. This exercises the relaxed validation against a real apiserver and protects the suspended -> unsuspended transition path used by Kueue. Refs: #3043 Signed-off-by: Sabari Narayana <[email protected]> * fix(api): block RuntimePatches change only when both states are unsuspended Per @mimowo's review, the previous condition (!oldSuspended) over-blocked the (oldSuspended=false, newSuspended=true) transition — i.e. suspending a running TrainJob while modifying spec.runtimePatches in the same call, which is safe because the JobSet activity check still guards against concurrent pod runs. Change the predicate to !oldSuspended && !newSuspended so the webhook only forbids RuntimePatches mutations when the TrainJob is staying fully unsuspended. Reword the error message accordingly so callers see explicitly that suspension at either side of the update is acceptable. Refs: #3043, kubeflow/trainer#3469 (review feedback) Signed-off-by: Sabari Narayana <[email protected]> * test: cover unsuspended->suspended atomic RuntimePatches transition Per @mimowo's request, complete the transition table coverage: - (false, false) -> blocked (existing test) - (true, true) -> allowed (existing test) - (true, false) -> allowed (added previously) - (false, true) -> allowed (added here) Mirrors the existing atomic-unsuspend cases — one unit case in jobset_test.go and one webhook integration entry in trainjob_test.go. Refs: #3043, kubeflow/trainer#3469 (review feedback) Signed-off-by: Sabari Narayana <[email protected]> * docs(api): shorten RuntimePatches validation comment Trim the comment block above checkRuntimePatchesImmutability to a 4-line summary, per @kaisoz's review feedback (the comment was longer than the change itself). Drops the kueue#8296 backreference and the k/k Job validation link — both are already in the PR description and commit history. Refs: kubeflow/trainer#3469 (review feedback) Signed-off-by: Sabari Narayana <[email protected]> * fix(api): use stdlib for suspend pointer checks in jobset webhook Per @tenzen-y's review nit, swap the ptr.Equal(x, ptr.To(true)) idiom for an inline nil-check + dereference. Equivalent semantics, no extra generic-comparison hop, and stays in the stdlib for these two lines. Refs: kubeflow/trainer#3469 (review feedback) Signed-off-by: Sabari Narayana <[email protected]> * test(jobset): cover (false,true) transition with active jobSet Per @tenzen-y's request, add a unit case where the old TrainJob is unsuspended, the new one is being suspended with a runtimePatches change, and the underlying JobSet still has an active replicatedJob. The relaxed suspend predicate alone would let this through; the JobSet activity check should still reject it. Refs: kubeflow/trainer#3469 (review feedback) Signed-off-by: Sabari Narayana <[email protected]> * test(integration): drop atomic transition cases duplicated by UTs Per @tenzen-y's review, the two atomic-transition ginkgo entries (unsuspend + RuntimePatches, suspend + RuntimePatches) duplicate the unit cases in jobset_test.go that already exercise the same validator path. The pre-existing "Should succeed to update runtimePatches when suspend is true" entry stays as the integration smoke test for the webhook chain. Refs: kubeflow/trainer#3469 (review feedback) Signed-off-by: Sabari Narayana <[email protected]> * fix(api): use new(true) for suspend pointer comparisons @tenzen-y is right that the value-form `new()` builtin is valid Go 1.26, which the repo already targets (`go 1.26.0` in go.mod). Apply his original suggestion verbatim — keeps `ptr.Equal` for the two-pointer comparison and replaces only `ptr.To(true)`, which is the smaller and more idiomatic delta vs the inline nil-check + deref pattern I tried in the previous commit. Refs: kubeflow/trainer#3469 (review feedback) Signed-off-by: Sabari Narayana <[email protected]> --------- Signed-off-by: Sabari Narayana <[email protected]>