Run:

    seq 100 | parallel -j10 -t --resume --results res/ 'seq 1024;sleep 1;seq 1024;true' & pid=$!; sleep 3; kill $pid
    seq 100 | parallel -j10 -t --resume --results res/ 'seq 1024;sleep 1;seq 1024;true' & pid=$!; sleep 3; kill $pid
    seq 100 | parallel -j10 -t --resume --results res/ 'seq 1024;sleep 1;seq 1024;true' & pid=$!; sleep 3; kill $pid
    seq 100 | parallel -j10 -t --resume --results res/ 'seq 1024;sleep 1;seq 1024;true' & pid=$!; sleep 3; kill $pid

The size of the output should be the same. But is not:

    4 -rw-rw-r-- 1 ota ota 4013 Apr 25 13:40 res/1/27/stdout
    4 -rw-rw-r-- 1 ota ota 4013 Apr 25 13:40 res/1/28/stdout
    4 -rw-rw-r-- 1 ota ota 4013 Apr 25 13:40 res/1/29/stdout
    4 -rw-rw-r-- 1 ota ota 4013 Apr 25 13:40 res/1/30/stdout
    8 -rw-rw-r-- 1 ota ota 8026 Apr 25 13:40 res/1/10/stdout
    8 -rw-rw-r-- 1 ota ota 8026 Apr 25 13:40 res/1/11/stdout
    8 -rw-rw-r-- 1 ota ota 8026 Apr 25 13:40 res/1/12/stdout
    8 -rw-rw-r-- 1 ota ota 8026 Apr 25 13:40 res/1/13/stdout
    8 -rw-rw-r-- 1 ota ota 8026 Apr 25 13:40 res/1/14/stdout

This is because GNU Parallel mistakenly assumes the job is done when the dir is created. What is needed is some way for GNU Parallel to determine that the job has finished.

One solution would be to add another file called 'joblog'. It could contain the full line from --joblog.

This way we can easily see if the job is incomplete (the file is missing) and also see if the job failed (Exitval != 0).

Thereby --resume, --resume-failed and --retry-failed can be implemented.