I am developing a generic package implementing a parallel algorithm and I want to provide generic IO routines for user types. I know that JLD2 may have breakage between Julia releases because the serialization of types is not guaranteed to be backwards-compatible. However, I would like to guarantee that if a user’s types are raw arrays then they will be able to read back in the stored data. Does JLD2 guarantee this? I can always allow a method for data storage that provides a custom serialization function to convert a user type to a raw array and then use JLD2 as backend to write that array.
AFAIK JLD2 does not uses Julia’s serialization, but a subset of HDF5, so it should be OK.
If by “raw arrays”, you mean commonly used primitive bitstypes like defined in Base like Int32 and Float64, you should be fine.
Yeah, that’s what I was getting at. How does JLD2 serialize arbitrary Julia types using only the HDF5 standard?
AFAIK it has its own format, but takes advantage of HDF5 features when it can.