Missing table properties about h5py HOT 4 CLOSED

From [email protected] on November 18, 2008 21:38:39
These properties exist in the current svn version; they'll be part of the upcoming
1.0 release. One issue with an "extendable" flag may be that not every axis is
guranteed to be extendable; if you declare the maxshape as (10,None,10), then only
the second axis can be extended. I'll see if there's a way to deal with this.

from h5py.

From [email protected] on November 19, 2008 13:48:47
Thanks! Regarding 'extendable', a developer will need to know the axises and
chunksizes to extend, so the 'None in t.maxsize' is about as good as we can get as a
single boolean flag. I could forsee uses like:

def extend(t,newchunks):
rank = len(t.shape)
if isinstance(newchunks,int):
newchunks = [newchunks]_rank
if len(newchunks) != rank:
raise ValueError('Invalid rank')
newsize = []
for nchunks,n,msize,csize in izip(newchunks,t.shape,t.maxsize,t.chunksize):
if msize is None:
raise ValueError('Cannot resize axis')
newsize.append( n+nchunks_csize )
t.resize( tuple(newsize) )

I'd rename the current 'extend' to 'resize', since it takes an absolute size and
rework 'extend' to do as above.

from h5py.

From [email protected] on November 19, 2008 15:39:22
It turns out you don't need to make the dataset size an exact multiple of the chunk
size; although HDF5 will allocate storage space in chunks you can specify any shape
you want. Dataset chunking is supposed to be transparent; I'm a little suspicious of
any method that wants you to do chunk arithmetic.

Extend() has been deprecated because it no longer hooks in to the HDF5 H5Dextend()
function, which actually did something slightly different (guaranteed a minimum
size). H5Dextend has coincidentally been deprecated in HDF5 1.8.

If you want to grow an existing array by a fixed amount you can do something like:

dset.resize([x+y for x, y in zip(dset.shape, extension_sequence)])

or for a single axis

dset.resize(dset.shape[ax]+amount, axis=ax)

from h5py.

From [email protected] on November 19, 2008 16:09:37
That simplifies everything. Never mind-- a plain resize works fine for me now that I
know that the chunking is a back end detail (to my great relief). I'd rather not
have to worry about growing by fixed amounts and simply resize arbitrarily as needed,
provided I don't change fixed dimensions.

from h5py.