kubeflow / trainer / 33132862846

69%

push

github

fix(framework): stop Flux reconcile panic on unset trainer fields (#3904)

spec.trainer and spec.trainer.numNodes are both optional and neither is
defaulted, so a TrainJob that omits them is admitted by the webhook and
then dereferenced unconditionally by the Flux plugin. controller-runtime
recovers the panic and requeues, so the TrainJob never gets a JobSet and
retries forever while the manager logs a stack trace each round.

The trainer PodSet count already carries the TrainJob's numNodes by the
time the component builders run, so the Flux entrypoints read the node
count from there rather than from the TrainJob.

Signed-off-by: rjgoyln <pt40419@gmail.com>

24 of 24 new or added lines in 1 file covered. (100.0%)

2696 of 3883 relevant lines covered (69.43%)

0.81 hits per line