netcdf4-python support for s3

[bnlawrence]

We believe that netcdf-c should support natural operations on a file on an S3 storage system. We need to test and demonstrate that we can do this with netcdf4-python (perhaps via a version with a specially compiled c library).

(We need this to avoid having to rebuild all of cf-python with h5netcdf with unknown implications for functionality and performance).

[valeriupredoi]

@bnlawrence @davidhassell @markgoddard looks like we are very, very close to an S3-abled netCDF4-python Unidata/netcdf4-python#1264

[valeriupredoi]

that is now possible via this biutiful thing done by my bud @zklaus conda-forge/libnetcdf-feedstock#178 - we'll have to wait for netCDF4-python folk to build new wheels and conda binaries, but that's now just a matter of time, and not a matter of "if" 😁

[zklaus]

Cheers 😄. Note that conda-forge/libnetcdf-feedstock#178 is not quite enough; we also need conda-forge/libnetcdf-feedstock#180.

[valeriupredoi]

Right on, cheers, Klaus! But the bigh heavy thing was that that you've done in 178, am assuming now it's just a matter of time, not of will and blah anymore 😁 I owes you a 🍺 bud

[valeriupredoi]

also pay attention to what Klaus mentions here conda-forge/libnetcdf-feedstock#178 (comment)

[valeriupredoi]

gentlemen @bnlawrence @davidhassell @markgoddard conda-forge/libnetcdf-feedstock#180 has now been merged, so in theory, we should be good to go to implement our use of netCDF4-python with S3 - also in light of #95 (comment) it looks like some mucking about will still be needed 😁 Of course, one needs to wait for wheels now ☸️

[markgoddard]

Something I noticed while working on byte order and missing data support - Kerchunk's SingleHdf5ToZarr uses h5py to read the netcdf metadata when creating a Zarr array. There definitely seems to be some oddness there in what we get back. In particular, when a byte order/endianness is specified on a variable, it always seems to show as big endian in the Zarr array. I have a workaround for this in #132, but I expect there are other issues, in particular around compressed data.

[valeriupredoi]

in S3-netCDF4 news there are bad news - S3 support had to be reverted from libnetcdf since it was causing flakiness apparently conda-forge/libnetcdf-feedstock#188 - I created a test branch to see when/how that's ready to work https://github.com/valeriupredoi/PyActiveStorage/tree/try_netcdf4

[valeriupredoi]

OK in light of the bad news listed above, I got some fairly solid progress done on this: it proves out we don't actually need S3-support in netCDF4 per se, all we have to do is be clever about how we read the file as a remote file:

    with fs.open(s3_testfile_uri, 'rb') as s3f:
        nc_bytes = s3f.read()

    ds = netCDF4.Dataset(f'inmemory.nc', memory=nc_bytes)

will read the S3 file into memory and dump it a very nice netCDF4 Dataset at the client: you can see it in action in this GA https://github.com/valeriupredoi/PyActiveStorage/actions/runs/6225016972/job/16894516376 - the problem is that calling netCDF4.Dataset on the memory object that'll dump all the data into a real and heavy netCDF4-like item. All we need to do is to understand how not should do that 😁

[bnlawrence]

I think the problem with that is that you've read the entire file across the network. Ideally we'd be using range-get and only getting metadata in the first instance, so it would remain lazy for actual data.

[zklaus]

You should be able to construct a simple https URL using bucket name and object name as https://localhost:9000/bucketname/object. Then you should be able to use #byte-range to open the whole thing without downloading it.

The caveats are:

• for this to work the policy needs to be set correctly on minio (see Public URL access not working minio/minio#7393)
• this negates part of the S3 advantage, since the URL must be public and might not offer the same resilience (particular wrt different hosts...)

But for this testing, that is appropriate. Hopefully, we'll resolve the S3 support before things move into production, pending the resolution of Unidata/netcdf-c#2739.

[valeriupredoi]

@bnlawrence exactly - I think there is a way to MacGyver that 😁 @zklaus cheers, but the problem is not solved by creating a public URL - we need to have support for s3:// types of file URLs + netCDF4python, but I'll try see what I can do with what we currently have. Hoping this will be sorted out upstream, like you did before (and that was so close!) 🍺

[markgoddard]

Something I noticed while working on byte order and missing data support - Kerchunk's SingleHdf5ToZarr uses h5py to read the netcdf metadata when creating a Zarr array. There definitely seems to be some oddness there in what we get back. In particular, when a byte order/endianness is specified on a variable, it always seems to show as big endian in the Zarr array. I have a workaround for this in #132, but I expect there are other issues, in particular around compressed data.

Just bumping this ^ While it may be an easy enough switch to netcdf4-python in PyActiveStorage when reading the metadata, swapping out h5py for netcdf4-python in kerchunk would be more difficult, and potentially more contentious.