That makes rather more sense. Since Tim seems to get a friendly error message in a situation that we believe to be similar to yours, I think a more comprehensive set of steps to reproduce--if indeed you can--will be needed to resolve any issues.
On Thursday, March 13, 2014 12:48:28 PM UTC-5, ben wrote: > > I meant composite! Grr. > > On Thursday, March 13, 2014 9:54:30 AM UTC-4, Patrick O'Leary wrote: >> >> On Thursday, March 13, 2014 8:48:30 AM UTC-5, Tim Holy wrote: >>> >>> IIRC it does actually store the definition of types in a "hidden" group >>> inside >>> the file. But in general it's hard to make use of: what if the type is >>> defined >>> inside a module that hasn't been loaded? So yes, HDF5/JLD are a little >>> bit >>> fragile when it comes to types; unfortunately, I don't think there is a >>> good >>> solution---it's basically a consequence of having modules (which are a >>> good >>> thing!). >>> >> >> The segfault is suboptimal though. Since you store the structure of the >> source type on serialization, can we check for the existence of an >> equivalently-defined type before deserializing? Though the fact that Ben >> said he changed the *abstract* supertype is confusing; why would that >> matter? Or do I misunderstand him? >> >
