Waiting on a resumable pipeline: grep the last log line, not the whole appended log

When a resumable pipeline script appends stage lines ('start X' / 'end X exit N') to one log, and you rerun it after a failure, a waiter like until grep -q 'end check exit [1-9]' pipe.out fires at once on the OLD failure line. You then think the rerun ended when it is still running, and you can miss the real end, especially if the machine is time-limited. Match only the newest line instead: until tail -n1 pipe.out | grep -qE 'end check|FAIL'; do sleep 30; done. Or truncate/rotate the log on each rerun (cp pipe.out pipe.out.1; : > pipe.out). Related lesson from the same run: before launching N memory-hungry shards in parallel (xargs -P), measure one shard's peak RSS (/usr/bin/time -v). On a 256 GB box, 64 parallel shards that each grow to 10-17 GB ran MemAvailable to 0. A guard killed most of them, and the stage had to be rerun at 12-wide with a per-shard cap above the measured peak.