PERF: refactor string construction benchmark

[ngoldbaum]

• closes #xxxx (Replace xxxx with the GitHub issue number)
• Tests added and passed if fixing a bug or adding a new feature
• All code checks passed.
• Added type annotations to new arguments/methods/functions.
• Added an entry in the latest doc/source/whatsnew/vX.X.X.rst file if fixing a bug or adding a new feature.

For asv peakmem benchmarks, the memory cost of the setup function is included in the benchmark result. The way this benchmark is currently written, the memory usage is dominated by constructing series_cat_arr, so the memory benchmark results aren't very useful for Series and DataFrame. I also added string[pyarrow] to the benchmark because why not?

The results before the refactor:

[ 8.33%] ··· strings.Construction.peakmem_cat_series_construction
                                ok
[ 8.33%] ··· ======== ======
              dtype         
             -------- ------
               str     329M 
              string   313M 
             ======== ======

[16.67%] ··· strings.Construction.peakmem_frame_construction
                                     ok
[16.67%] ··· ======== ======
              dtype         
             -------- ------
               str     313M 
              string   313M 
             ======== ======

[25.00%] ··· strings.Construction.peakmem_series_construction
                                    ok
[25.00%] ··· ======== ======
              dtype         
             -------- ------
               str     313M 
              string   313M 
             ======== ======

[33.33%] ··· strings.Construction.time_cat_series_construction
                                   ok
[33.33%] ··· ======== ==========
              dtype             
             -------- ----------
               str     213±0ms  
              string   43.1±0ms 
             ======== ==========

[41.67%] ··· strings.Construction.time_frame_construction
                                        ok
[41.67%] ··· ======== ==========
              dtype             
             -------- ----------
               str     37.9±0ms 
              string   54.5±0ms 
             ======== ==========

[50.00%] ··· strings.Construction.time_series_construction
                                       ok
[50.00%] ··· ======== ==========
              dtype             
             -------- ----------
               str     38.1±0ms 
              string   39.5±0ms 
             ======== ==========

And after:

[25.00%] ··· strings.Construction.peakmem_construction
                                           ok
[25.00%] ··· ==================== ====== ================ =================
             --                                     dtype                  
             -------------------- -----------------------------------------
                   pd_type         str    string[python]   string[pyarrow] 
             ==================== ====== ================ =================
                    series         310M        310M              316M      
                    frame          310M        310M              316M      
              categorical_series   327M        313M              319M      
             ==================== ====== ================ =================

[50.00%] ··· strings.Construction.time_construction
                                              ok
[50.00%] ··· ==================== ========== ================ =================
             --                                       dtype                    
             -------------------- ---------------------------------------------
                   pd_type           str      string[python]   string[pyarrow] 
             ==================== ========== ================ =================
                    series         39.3±0ms      37.6±0ms          48.9±0ms    
                    frame          40.0±0ms      41.7±0ms          56.4±0ms    
              categorical_series   215±0ms       44.7±0ms          57.0±0ms    
             ==================== ========== ================ =================

Note the different, lower, peak memory usages. I also find it a bit easier to compare results as two parameterized benchmarks.

[MarcoGorelli]

Thanks @ngoldbaum for your PR!

Agree that this is much easier to read, good one

Could you please explain why now we need to pass

dtype=self.dtype_mapping[dtype]

? Looks like the default is object, so here this only makes a difference for str dtype - why is it better to pass dtype=str for that one?

[MarcoGorelli]

nit: elif

[ngoldbaum]

In principle we don’t but IMO it makes the timing benchmarks more fair because it avoids a copy to numpy’s string dtype inside pandas. I have a version of this benchmark with the new string dtype I’ve been working on and avoiding that copy is substantially faster there.

[MarcoGorelli]

Looks good, thanks @ngoldbaum !

Leaving open a bit in case others have comments

[mroeschke]

Thanks @ngoldbaum