Add `__orthogonal_indexing__` flag

[davidhassell]

Some libraries include an __orthogonal_indexing__ property to indicate to the user whether or not a variable utilises orthogonal indexing. For instance, netCDF4, cfdm, and cf-python implements this API, and it's very useful because it tells downstream applications whether or not they themselves need to implement orthogonal indexing.

For these other libraries just mentioned, __orthogonal_indexing__ is always True, but for a pyfive Dataset, orthogonal indexing is dependent on the data layout - contiguous data does not support it, but chunked data does.

Therefore we could add an __orthogonal_indexing__ property that returns True or False depending on the data layout.

[davidhassell]

PR to follow that will be built on #170, so I'll only post it when that is resolved.

[bnlawrence]

So, we do have the option for pseudo-chunking for contiguous data, which relies on orthogonal indexing. But why doesn't orthogonal indexing apply to contiguous data using the memory map feature?

[davidhassell]

So, we do have the option for pseudo-chunking for contiguous data, which relies on orthogonal indexing

Ah ha! I thought that might have been the case. How does the user choose this?

But why doesn't orthogonal indexing apply to contiguous data using the memory map feature?

That's because numpy doesn't do orthogonal indexing.

[bnlawrence]

The option def _get_direct_from_contiguous(self, args=None): uses psuedo chunking instead of memory mapping.

[davidhassell]

OK - I see. The new code I have is currently:

class Dataset(object):

<snip>

    @property
    def __orthogonal_indexing__(self):
        """Flag to indicate whether indexing is orthogonal.

        """
        return self.id.chunks is not None

We could easily put the contiguous logic into def __orthogonal_indexer__:

    @property
    def __orthogonal_indexing__(self):
        """Flag to indicate whether indexing is orthogonal.

        """
        if self.id.chunks is not None:
            # Chunked data
            return True

        if not (isinstance(self.id._ptype, (P5ReferenceType, P5VlenStringType, P5SequenceType)) and not self.id.posix:
            # Contiguous data accessed with pseudo chunking
            return True

        # All other cases
        return False