Batch jobs fail differently from user-facing services. Nobody is watching, so there's no immediate signal. Failures surface hours later and somewhere else, usually as a missing report, an incomplete dataset, or a downstream job that started on schedule with nothing to process.

That delay is the real cost. By the time anyone notices, the window to rerun cleanly may have closed, and you're reconciling partial output against a deadline. Compute environments that lose an instance mid-job can leave work in a state that's harder to recover from than an outright failure would have been.

Testing your compute environment tells you what your jobs do when an instance disappears partway through: whether they resume, restart cleanly, or quietly produce incomplete results. The last outcome is the dangerous one, and it's the one you're least likely to catch without deliberately looking for it.