Concurrent access to PSpecContainer #193
Conversation
…allow concurrent access
ghost
assigned 
There was a problem hiding this comment.
This looks like a great addition to the pspec container object. Most of my comments are minor, and are mostly stylistic.
I did have a couple of larger questions about how this application works within the SWMR paradigm. According to the SWMR user guide, the writer is only permitted to append data along an unlimited dimension, or modify existing datasets. It is explicitly not allowed to add new groups, datasets, attributes, etc. I am not very familiar with the typical access patterns of the pspec container object, but it seems like some object methods like set_pspec can create new groups and datasets. Clearly there are tests that are passing, so perhaps the SWMR feature is being used in such a way that it's compatible with the limitations of the HDF5 library. However, I am nervous about using the features in undocumented ways, which may lead to headaches down the road.
Additionally, use of the SWMR feature requires that the underlying file system be POSIX-compliant. A notable exception to this is NFS or other network-mounted filesystems. In the documentation of this parallel access feature on hera_pspec, the user should be informed of this requirement, so that it can be used appropriately.
Add a 'transactional' mode and enable HDF5's "single writer, multiple reader" (SWMR) mode to allow concurrent access to
PSpecContainerobjects.The transactional mode is enabled when the container is instantiated with the
keep_open=Falsekwarg. This keeps the HDF5 file closed except for when actions are being performed on it, when it is briefly opened and then closed again. This is helpful if multiple processes need to write out to the same file for example, as HDF5 files cannot be opened in read-write mode by more than one process.The SWMR mode allows multiple processes to open the file for reading at the same time, as long as at most one process has opened it in read-write mode. Generally, the read-write process should open the file first; if a read-only version has been opened beforehand, opening in read-write mode may fail.