Performance Shift(s): ca42c30b

[github-actions]

Benchmark comparison has identified performance shifts at

• commit ca42c30 ([CI Bot] environment lockfiles auto-update #5163).

Please review the report below and take corrective/congratulatory action as appropriate 🙂

Performance shift report

       before           after         ratio
     [504c188f]       [ca42c30b]
     <main~2>         <main~1>  
+        407±20μs         517±20μs     1.27  coords.AuxCoordLazy.time_bounds
+        391±20μs         510±20μs     1.30  coords.AuxCoordLazy.time_points
+        411±20μs         520±30μs     1.26  experimental.ugrid.ConnectivityLazy.time_indices(1000000)
+        409±20μs         515±20μs     1.26  experimental.ugrid.ConnectivityLazy.time_indices(6)

Generated by GHA run 4227658143

[trexfeathers]

Presumably linked to the new Dask version? Are we OK with this? Less than 1 millisecond, even for highly scaled examples (e.g. time_indices(1000000)).

[lbdreyer]

Presumably linked to the new Dask version? Are we OK with this? Less than 1 millisecond, even for highly scaled examples (e.g. time_indices(1000000)).

It is interesting that it doesn't scale with the number of points. It implies that perhaps it is something to do with the setting up of the lazy array (with dask as you say) and the only reason these benchmarks caught it is because they are so fast?

If that were the case it seems like this change is very minimal and not to be worried about

[pp-mo]

@lbdreyer doesn't scale with the number of points.

From my reading, it is not actually actually reading any data since the test example is made from a numpy array.
So it is like da.from_array(np.array(...)).compute(), which simply returns the underlying array object. No copying or moving data is involved.

>>> import numpy as np
>>> import dask.array as da
>>> np1 = np.arange(24.).reshape((4,3,2))
>>> da1 = da.from_array(np1)
>>> da1.compute() is np1
True
>>>