Custom-container operators bypass operator-runner's ResourceSampler, so they
emit no resourceSeries in result.json — the platform resource monitor showed
GPU=0 even while nvidia-smi in the pod reported 44-91% utilization.
The launcher now samples GPU usage itself for custom-container pods (nvidia-smi
exec into the pod every 10s while Running) and executeNode falls back to those
platform-collected samples when the operator did not self-report. Builtin
operator-runner pods still self-report a richer series (CPU+memory too) and are
skipped to avoid duplicate work.
- launcher.go: store rest.Config; background GPU sampler goroutine gated on
custom image (Image != OperatorRunnerImage); nvidia-smi exec + csv parse
mirroring operator-runner's resource_monitor.py rollup (max util, summed
memory, deduped names)
- executor.go: pass OperatorRunnerImage; map launcher GPUSample -> domain
ResourceSample when parsed resourceSeries is empty
- droboticflow.template.yaml: grant worker Role pods/exec (nvidia-smi sampling)
- launcher_gpu_sampler_test.go: parser unit tests
- go.mod/go.sum: remotecommand transitive deps; go mod tidy also corrects
bce-sdk-go/x/crypto to direct (was mislabeled indirect)
Degrades gracefully: no restConfig / no nvidia-smi in image / no pods/exec RBAC
/ CPU-only pod -> zero samples, no regression vs. current behavior.
Co-Authored-By: Claude <noreply@anthropic.com>
Include the accumulated workflow scheduling, run history, subtree rerun, operator UI, deployment, data asset, object storage, database browser, and documentation changes currently in the worktree.