merge consecutive reads in shards

[jstriebel]

This PR adds a feature for the v3 sharding implementation. When multiple chunks are read at once with a store that supports partial reads, each chunk was requested separately before, even if it was in the same shard. This PR adds the feature to merge consecutive (on a byte-level) requests in the same shard.

TODO:

• Add unit tests and/or doctests in docstrings
• New/modified features documented in docs/tutorial.rst
• Changes documented in docs/release.rst
• GitHub Actions have all passed
• Test coverage is 100% (Codecov passes)

[codecov]

Codecov Report

Merging #1358 (a98c7b8) into main (5ece3e6) will decrease coverage by 0.04%.
The diff coverage is 83.33%.

@@             Coverage Diff             @@
##              main    #1358      +/-   ##
===========================================
- Coverage   100.00%   99.96%   -0.04%     
===========================================
  Files           36       36              
  Lines        14775    14804      +29     
===========================================
+ Hits         14775    14799      +24     
- Misses           0        5       +5     

Impacted Files	Coverage Δ
zarr/_storage/v3_storage_transformers.py	97.97% <83.33%> (-2.03%)	⬇️

[rabernat]

Thanks for working on this important optimization! @martindurant implemented something similar at the fsspec level in fsspec/filesystem_spec#1063. I'm curious how this implementation here would interact with that one. Where in this stack should this optimization live?

[normanrz]

Good point. We weren't aware of that optimization. I almost think it makes more sense in the fsspec. @jstriebel wdyt?

[rabernat]

I almost think it makes more sense in the fsspec.

That's a reasonable point of view. However, currently not all zarr I/O goes through fsspec. This is a rather major architectural issue that we need to confront for zarr-python: how dependent do we want to be on fsspec? If we require fsspec as a core dependency, we could remove quite a bit of code from zarr-python itself. Currently, we're somewhere in the middle--we rely on fsspec for some things, but we still have many of our own implementations for different stores.

We need to discuss this as a team.

[martindurant]

Totally agree with @rabernat , of course you might know which way I lean.

For merging the requests via fsspec, this should be available via the cat_ranges() method - it is intended but not implemented yet.

[jhamman]

closing since now that the sharding implementation is in a codec, this would need to be reworked. This feature conversation is also happening as part of #1583