improve column access speed for sparse matrices (#742)

* improve column access speed for sparse matrices

* add FAQ entry about data format performance

* add note about using --sparse flag for prepare command

* clean up for PR review

* Update docs/faq.md

Co-Authored-By: bkmartinjr <bruce@chanzuckerberg.com>

* improvements to big data faq

docs/faq.md

@@ -40,6 +40,14 @@ Currently this is not supported directly, but you should be able to do this your
 - `.X` is used to display expression (histograms, scatterplot & colorscale) and to compute differential expression
 - `.obsm` is used for layout
 
+#### I have a BIG dataset - how can I make cellxgene run as fast as possible?
+
+If your dataset requires gigabytes of disk space, you may need to select an appropriate storage format in order to effectively utilize `cellxgene`. Tips and tricks:
+
+- `cellxgene` is optimized for columnar data access. For large datasets, format the expression matrix (`.X`) as either a [SciPy CSC sparse matrix](https://docs.scipy.org/doc/scipy/reference/generated/scipy.sparse.csc_matrix.html) or a dense Numpy array (whichever creates a smaller `h5ad` file). If you are using `cellxgene prepare`, include the `--sparse` flag to ensure `.X` is formatted as a CSC sparse matrix (by default, `.X` will be a dense matrix).
+- `cellxgene` start time is directly proportional to `h5ad` file size and the speed of your file system. Expect that large (eg, million cell) datasets will take minutes to load, even on relatively fast computers with a high performance local hard drive. Once loaded, exploring metadata should still be quick.
+- If your dataset size exceeds the size of memory (RAM) on the host computer, differential expression calculations will be extremely slow (or fail, if you run out of virtual memory).
+
 # Algorithms
 
 #### How are you computing and sorting differential expression results?

server/app/scanpy_engine/scanpy_engine.py

@@ -3,6 +3,7 @@ import warnings
 import numpy as np
 from pandas.core.dtypes.dtypes import CategoricalDtype
 import scanpy as sc
+from scipy import sparse
 
 from server.app.driver.driver import CXGDriver
 from server.app.util.constants import Axis, DEFAULT_TOP_N
@@ -291,6 +292,22 @@ class ScanpyEngine(CXGDriver):
             df = df[fields]
         return encode_matrix_fbs(df, col_idx=df.columns)
 
+    @staticmethod
+    def slice_columns(X, var_mask):
+        """
+        Slice columns from the matrix X, as specified by the mask
+        Semantically equivalent to X[:, var_mask], but handles sparse
+        matrices in a more performant manner.
+        """
+        if var_mask is None:    # noop
+            return X
+        if sparse.issparse(X):  # use tuned getcol/hstack for performance
+            indices = np.nonzero(var_mask)[0]
+            cols = [X.getcol(i) for i in indices]
+            return sparse.hstack(cols)
+        else:   # else, just use standard slicing, which is fine for dense arrays
+            return X[:, var_mask]
+
     @requires_data
     def data_frame_to_fbs_matrix(self, filter, axis):
         """
@@ -313,9 +330,7 @@ class ScanpyEngine(CXGDriver):
             raise FilterError("filtering on obs unsupported")
 
         # Currently only handles VAR dimension
-        X = self.data._X
-        if var_selector is not None:
-            X = X[:, var_selector]
+        X = self.slice_columns(self.data._X, var_selector)
         return encode_matrix_fbs(X, col_idx=np.nonzero(var_selector)[0], row_idx=None)
 
     @requires_data