kubeflow / trainer / 29758781335

68%

push

github

fix(runtimes): merge resourcesPerNode with Runtime container resources (#3602)

* fix(controller): merge resourcesPerNode with Runtime container resources

When a TrainJob sets resourcesPerNode, merge it with the referenced
Runtime's node container resources instead of replacing the entire block.
TrainJob keys override matching keys; unset keys inherit from the Runtime.

This fixes the case where specifying only GPU silently drops CPU/memory
from the Runtime and pushes pods to BestEffort QoS.

Fixes kubeflow/trainer#3551

Signed-off-by: Sridhar1030 <sridharpillai75@gmail.com>

* fix(controller): address review feedback for resourcesPerNode merge

Signed-off-by: Sridhar1030 <sridharpillai75@gmail.com>

* fix(runtimes): use SMP for claims merge and deduplicate apply path

- Replace manual merge with Strategic Merge Patch so Claims are merged
  by name (patchMergeKey:name), consistent with Limits and Requests
- Move resource merge before FromTypedObjWithFields to keep typed spec
  and apply configuration in sync, removing duplicated inline patching
- Extract MergeResourceRequirements into pkg/runtime for isolated testing
- Add merged_resourcesPerNode case to TestTrainingRuntimeNewObjects

Signed-off-by: Sridhar1030 <sridharpillai75@gmail.com>

* fix(runtimes): move MergeResourceRequirements to util and drop unused helpers

Signed-off-by: Sridhar1030 <sridharpillai75@gmail.com>

* fold MergeResourceRequirements into trainingruntime.go

Signed-off-by: Sridhar1030 <sridharpillai75@gmail.com>

* fix(runtimes): fold resource merge into ReplicatedJobs loop

Signed-off-by: Sridhar1030 <sridharpillai75@gmail.com>

* fix(runtimes): fold resource merge into ReplicatedJobs loop

Signed-off-by: Sridhar1030 <sridharpillai75@gmail.com>

---------

Signed-off-by: Sridhar1030 <sridharpillai75@gmail.com>

40 of 53 new or added lines in 2 files covered. (75.47%)

2548 of 3733 relevant lines covered (68.26%)

0.8 hits per line