Sogo Mineo writes in HSC-1414:

See the following lines:

meas_algorithms/HSC-4.0.0/python/lsst/meas/algorithms/objectSizeStarSelector.py:466
in ObjectSizeStarSelector.selectStars():

if psfCandidate.getWidth() == 0:

psfCandidate.setBorderWidth(self._borderWidth)

psfCandidate.setWidth(self._kernelSize + 2*self._borderWidth)

psfCandidate.setHeight(self._kernelSize + 2*self._borderWidth)

In reduceFrames, these lines set the width of psfCandidate to be 21
for the first time the execution reaches there.

When the first CCD image has been processed, the worker process
continues to process another CCD image, and the execution reaches
here again. This time, psfCandidate.getWidth() is 41, because
psfexPsfDeterminer has set it to be 41, and the value has been
retained because the width is a static member. And so, for the second
CCD image, the width of psfCandidate is not 21 but 41.

Since psfCandidates are widened, stars positioned at edges of images
are rejected. It results in a smaller number of PSF candidates than expected.

Only CCD images that are initially given to the worker processes
are processed with psfCandidate.getWidth() == 21. The other CCD images are
processed with psfCandidate.getWidth() == 41.
When the number of SMP cores changes, CCD images are processed with different
parameters.

The change in the number of PSF candidates results in different Psf, a different
result of image repair, and different catalogs.

The line numbers are different on the LSST side because of refactoring (objectSizeStarSelector.py:466 has moved to starSelector.py:148), but the bug is still present. The main problem appears to be that the PsfCandidate elements are static, are being set in both the star selector and the PSF determiner and one of those is conditional on what the value is. I will investigate moving the static class variables to instance variables — the desired size appears to vary by context, so it shouldn't be a class variable.