Performance Shift(s): `14cf56ab`

[github-actions]

Benchmark comparison has identified performance shifts at

• commit 14cf56a (Upgrade iris.util.mask_cube #4889).

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

Performance shift report

       before           after         ratio
     [7a5e68d5]       [14cf56ab]
     <main~1>         <main>    
+     5.34±0.05μs      6.64±0.07μs     1.24  cube.AuxCoord.time_return_coords
+     5.48±0.05μs      6.88±0.05μs     1.26  experimental.ugrid.Mesh.time_connectivities(1000000)
+     5.47±0.07μs      6.79±0.04μs     1.24  experimental.ugrid.Mesh.time_connectivities(6)
+     6.70±0.06μs      8.26±0.04μs     1.23  experimental.ugrid.Mesh.time_coords(1000000)
+     6.72±0.04μs      8.21±0.05μs     1.22  experimental.ugrid.Mesh.time_coords(6)
+     5.30±0.04μs      6.69±0.06μs     1.26  experimental.ugrid.Mesh.time_remove_connectivities(1000000)
+     5.32±0.05μs      6.68±0.04μs     1.26  experimental.ugrid.Mesh.time_remove_connectivities(6)
+        22.9±1μs         28.9±2μs     1.26  experimental.ugrid.MeshLazy.time_connectivities(1000000)
+      19.8±0.9μs         26.3±2μs     1.33  experimental.ugrid.MeshLazy.time_connectivities(6)
+        28.6±2μs         34.5±2μs     1.21  experimental.ugrid.MeshLazy.time_coords(1000000)
+        26.0±1μs         34.9±3μs     1.34  experimental.ugrid.MeshLazy.time_coords(6)
+        24.5±1μs         31.4±2μs     1.28  experimental.ugrid.MeshLazy.time_remove_connectivities(1000000)
+        22.6±1μs         30.1±3μs     1.33  experimental.ugrid.MeshLazy.time_remove_connectivities(6)
+        43.2±2μs         55.2±3μs     1.28  experimental.ugrid.MeshLazy.time_remove_coords(6)

Generated by GHA run 2828850734

[rcomer]

Ay caramba!

[rcomer]

As far as I can tell from a quick grep, iris.util.mask_cube is not used anywhere in Iris except the tests. So I don't think this can be real. @trexfeathers do you agree?

[trexfeathers]

As far as I can tell from a quick grep, iris.util.mask_cube is not used anywhere in Iris except the tests. So I don't think this can be real. @trexfeathers do you agree?

I'm almost certain this IS real. If it were noise it wouldn't appear in such closely aligned benchmarks, nor have such closely aligned ratios.

IIRC you made some minor changes elsewhere in that PR too?

[trexfeathers]

@rcomer and I have proven that this is caused by the introduction of a lazy import in common.metadata.metadata_filter. This is consistent with why no there is no performance scaling with Coord/Connectivity size, and also means there should be no scaling with volume of Coords/Connectivitys. So there is no reason to suspect that this microsecond regression could indicate larger regressions for real world use cases. Closing.

[rcomer]

Thanks for the benchmarking tutorial @trexfeathers!