Task retries are triggered only if the error raised from the compute call is a QCEngine RandomError. The Psi4 harness has a method to try and classify some failure modes, with general random failures with no error message given a new generic error message which does not contain the required keywords to trigger a retry. This is problematic in procedures like torsiondrives or optimizations as the entire calculation is stopped when in practice they normally work fine after retrying the single point calculation which has failed.
Seems to happen when running a pool of workers executing psi4 single points on the same computer, see attached qcschema temp file which just ends suddenly when computing the gradient.
Describe the bug
Task retries are triggered only if the error raised from the compute call is a QCEngine
RandomError
. The Psi4 harness has a method to try and classify some failure modes, with general random failures with no error message given a new generic error message which does not contain the required keywords to trigger a retry. This is problematic in procedures like torsiondrives or optimizations as the entire calculation is stopped when in practice they normally work fine after retrying the single point calculation which has failed.I think simply changing the error message to include the correct keyword (segmentation fault) would fix this.
To Reproduce
Seems to happen when running a pool of workers executing psi4 single points on the same computer, see attached qcschema temp file which just ends suddenly when computing the gradient.
qcschema.txt
Expected behavior
A psi4 task which fails with no error message should be restarted.
Additional context