Open gipert opened 2 weeks ago
Oh, I see the problem. read
is being passed an index mask as a boolean array, which I don't think is supposed to be supported. It worked in the old version because it somehow passed all the checks before actually applying the mask, but once the mask got applied h5py's high level interface handled it. With the low level interface, this would have to be handled explicity (actually the easier solution is probably just to convert the mask into an entry list).
@ggmarshall should this be changed on the pargen side or should make pydataobj support it?
This PR has been accepted, but I'll leave the conversation "open" until we confirm that this is a feature we want to keep around long term (otherwise we will have to fix this in the dataflow and then revert the change)
Apparent regression seen with v1.10 at NERSC, compared to v1.7:
Have no time to debug further, but downgrading to v1.7 seems to solve the issue.