Diagnose a troubled cluster
Use these three fictional incidents to practice choosing evidence. Do not jump straight to a restart.
Match the symptom to the investigation
Inspect input delivery and preparation
Compare its communication path
Inspect shared loading demand
Three different symptoms
Idle GPUs after a dataset change: storage latency rose and the input now contains many small files. Compare read and preparation time before blaming calculation.
One node slows the whole job: its standalone math test passes, but communication performs worse. Compare its links, placement, and configuration with a healthy peer.
Recovery hurts healthy users: every returning replica loads its model together and storage latency spikes. Investigate the shared recovery demand.
Write an investigation
For each case, note impact, recent changes, one leading hypothesis, and evidence that could disprove it.
Then choose a safe containment step and a recovery check. Disruptive tests need authorization; physical faults may need qualified facilities staff.
Check your answer
The three starting points are input delivery, the affected communication path, and synchronized recovery load. These are hypotheses, not proven causes.
A useful answer narrows the problem and verifies completed work afterward. The same blind restart for all three cases does neither.