Dask PCA support

[selmanozleyen]

Hi,

Just wanted to start the PR. Passes the tests except one. Also need to deal with solver names since they don't correspond to anything dask uses.

Also refactored where the DaskArray mock class is.

Pinging @ivirshup

[codecov]

Codecov Report

Merging #2563 (3d86e82) into master (b3d5a6c) will increase coverage by 0.04%.
The diff coverage is 89.13%.

❗ Current head 3d86e82 differs from pull request most recent head 8541ce3. Consider uploading reports for the commit 8541ce3 to get more accurate results

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #2563      +/-   ##
==========================================
+ Coverage   72.27%   72.32%   +0.04%     
==========================================
  Files         105      105              
  Lines       11753    11785      +32     
==========================================
+ Hits         8495     8524      +29     
- Misses       3258     3261       +3     

Files Changed	Coverage Δ
scanpy/preprocessing/_normalization.py	88.88% <ø> (ø)
scanpy/testing/_pytest/marks.py	84.61% <ø> (ø)
scanpy/preprocessing/_pca.py	94.30% <89.13%> (-0.94%)	⬇️

[flying-sheep]

looks straightforward enough!

1. please make sure the docs look right and you are using links. You probably have to add a line to intersphinx_mapping in docs/conf.py to be able to link to the dask-ml docs.
2. the logic seems straightforward, but the user should know when which svd_solver argument works, as that part is getting complicated.
3. you have a lot of duplicated lines that can just be moved to after their if/else block to deduplicate them

[flying-sheep]

scanpy/scanpy/preprocessing/_pca.py

Lines 191 to 192 in 3c76af9

	if not zero_center or random_state or svd_solver != 'arpack':
	logg.debug('Ignoring zero_center, random_state, svd_solver')

at this point, svd_solver is None, right? Please make this line work with both None and the explicit value

[selmanozleyen]

@flying-sheep , haven't considered all combinations yet but wanted to check if this way is good for testing warnings.

[flying-sheep]

Looks great! You can get rid of expect_warning, as you can just simply check if expected_warning_message is not None instead, otherwise pretty ideal!

[flying-sheep]

You also need to fix the links. dask.arrays.linalg.svd_compressed should be dask.arrays.linalg.svd_compressed, then the docs will build.

[flying-sheep]

great! once the docs build (and you’re happy with it), that’s ready