This fixes a reference to a global variable instead of self.

It happened to work because the published did have a global variable holding
that instance.

This defines some exceptions to be used in publisher root classes so as to
have consistent behav...

This adds publisher root class registration and some docstrings.

By trying an ad hoc method of each registered root publisher class.

So that the collection of such classes can be queried.

The docstring renders them unnecessary.

Use publisher root classes

With no checks other than those in the original code (e.g. specifying a plain
file will fail str...

Consistent with the existing helper that only checks for filesystem paths.

The relative path is converted to the root's representation (also for
sane/canonical representat...

Define it for the dir publisher root to pull watching files out of main
publisher code.

For non-directory publisher roots, it may make no sense to have a path that
crosses from the rea...

However, uses of that may be an indication that more changes are needed there
to decouple code f...

Including internal check for valid dataset relative path.

And start using that at the publisher.

With some potential abstract methods to define its interface.

Carry compression params on import/export.

This updates `cat2export` and (especially) `cat2im...

Using the same technique as when exporting arrays.

The technique to keep memory usage low and not create an array each time is
similar to import co...

They are `Iterator` (they can be iterated directly), not `Iterable`.

After creating an empty dataset.

No need for it to be that nested.

Also for similarity with export code.

They are redundant now, but they will be needed when an empty dataset is
created, so include the...

For similarity with import code. Getting the creator does not perform any
changes, but it still...

For similarity with import code.

For similarity with import code.

Since they are only used for that. Also, remove related, obsolete todo
comment.

For consistency.

Especially for the case of a 1D array where the blockshape was set, since no
chunkshape is store...

For Blosc2 arrays, frames and compressed files (plain files still use the
default).

If other filters (like HDF5 shuffle) are enabled, the schunk may be
scrambled (and there is no w...

Since nothing has been created yet on disc.

Otherwise, getting `(array, schunk)` was still prone to segmentation faults if
the GC decided to...

This hides Blosc2 arguments from the caller, and storage issues from Blosc2
conversion machinery...

The empty array already had chunks which were pushed back by the insertion
operation, so they we...

Source data is still slurped and re-compressed into memory to create the
chunks to be stored.

Add `cat2export tool` (Caterva to HDF5).

This adds the `caterva2.tools.cat_to_hdf5` a.k.a. `c...

Missing example subsection.

Undo root being part of dataset path in clients

So that exported dataset enjoy some compression instead of being stored
completely decompressed....

Since it depends on the kind of dataset exported.