Launcher should only re-run unsuccessful runs

[prabhuramachandran]

Launcher currently re-launches everything when called again. Instead Launcher should only run previous runs if the earlier invocation failed (with a non-zero exit status) or explicitly asked to force-rerun. This would also make it easy to change the parameters and add additional parameters ranges to explore without having to worry about re-running everything.

[melver]

Similar to feature request #9, this sounds useful, but is not realizable for other instances of Launcher, e.g. QLauncher.

The exit-code is also not a one-size-fits-all solution, as there are cases where a range of exit-codes, not just zero, are valid (e.g. some of my experiments have this property). A more generic solution would be to have an application-specific way to detect failure of a previous invocation, and then either apply a filter or the proposed subtraction #5. Finally, the way your process the results would then be able to merge multiple invocations' results.

[jlstevens]

I agree that this feature request is potentially quite thorny and problematic. In particular, exit codes won't work for QLauncher as melver points out so you would need some other way of finding out whether a job failed or not.

I also worry about the idea of merging multiple invocations due to version control reasons but isn't this already possible by disabling the timestamp?

@melver Isn't that what you do yourself? Don't you ever re-launch some jobs with the same parameters as used in a previous launch e.g to fill in jobs that might have mysteriously failed? Then won't the results accumulate in the same directory as a previous launch?

[melver]

On Sun, 10 May 2015, J.L Stevens wrote: @melver Isn't that what you do yourself? Don't you ever re-launch some jobs with the same parameters as launched previously e.g to fill in jobs that might have mysteriously failed and have the results accumulate in the same directory as a previous launch?

I don't, never disabled timestamps.

[jlstevens]

I don't, never disabled timestamps.

Hmm. Maybe not any more but I am sure you requested this at some point (possibly in the distant past). Should check that this still works (maybe even write some tests for it).