ASV: benchmark for DataFrame.Update

[aureliobarbosa]

• 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.

@rhshadrach and @datapythonista

Benchmarks for the DataFrame.Update method are implemented, as suggested on PR #57637 by @datapythonista and based on the manual tests done by @rhshadrach on #55634. By running ASV locally we measure the change in the performance due to the fix on #57637, as indicated below:

Change	Main (2.2.1) [8da8b54]	PR (3.0) [93ea131]	Ratio	Benchmark (Parameter)
-	1.13±0.02ms	887±10μs	0.78	frame_methods.Update.time_to_update_random_indices
-	1.19±0.02ms	840±9μs	0.71	frame_methods.Update.time_to_update_small_frame_big_arg
-	1.27±0.01ms	859±9μs	0.67	frame_methods.Update.time_to_update_big_frame_small_arg

Regards

[rhshadrach]

Thanks for the PR! This looks good, except the runtime for the benchmark is a bit low (can make it hard to differentiate regressions vs noise). Can you increase the parameters to get runtime up to ~10ms.

[aureliobarbosa]

Can you increase the parameters to get runtime up to ~10ms.

@rhshadrach Think I am done here.

Data was increased twofold, from $10^4$ to $10^6$ points. Each test is now taking ~4ms. Didn't check the scaling, but it seems to be enough because the data size is similar to other benchmarks in the frame_methods.py. If needed, I can increase the data again.

Change	Before [4fe49b1]	After [93ea131] <fix_gh55509_v2~14>	Ratio	Benchmark (Parameter)
-	4.47±0.01ms	3.79±0.1ms	0.85	frame_methods.Update.time_to_update_big_frame_small_arg
-	4.31±0.03ms	3.06±0.03ms	0.71	frame_methods.Update.time_to_update_random_indices
-	2.94±0.02ms	1.10±0.01ms	0.37	frame_methods.Update.time_to_update_small_frame_big_arg

It would be interesting to have some guidance on the best data size for a useful test, but that would be a whole work on itself.

EDIT: Table

[rhshadrach]

Suggested change

	rng = np.random.default_rng()
	self.df = DataFrame(rng.uniform(size=(100_000, 10)))
	idx = rng.choice(range(100_000), size=100_000, replace=False)
	self.df_random = DataFrame(self.df, index=idx)
	idx = rng.choice(range(100_000), size=10_000, replace=False)
	cols = rng.choice(range(10), size=2, replace=False)
	self.df_sample = DataFrame(
	rng.uniform(size=(10_000, 2)), index=idx, columns=cols
	)
	rng = np.random.default_rng()
	self.df = DataFrame(rng.uniform(size=(1_000_000, 10)))
	idx = rng.choice(range(1_000_000), size=1_000_000, replace=False)
	self.df_random = DataFrame(self.df, index=idx)
	idx = rng.choice(range(1_000_000), size=100_000, replace=False)
	cols = rng.choice(range(10), size=2, replace=False)
	self.df_sample = DataFrame(
	rng.uniform(size=(100_000, 2)), index=idx, columns=cols
	)

Can you go one more order of magnitude larger. You increased by two orders of magnitude and saw very little change in runtime - meaning that you were mostly measuring overhead. With the above changes, I'm seeing:

%timeit df.update(df_sample)
%timeit df_random.update(df_sample)
%timeit df_sample.update(df)
33.9 ms ± 220 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
46.3 ms ± 744 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
19.4 ms ± 268 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

which is more in the realm of what we want for ASVs.

[aureliobarbosa]

Done.

Just for consistency, the benchmarks comparing the last version of DataFrame.Update() and the current one. At some point there is some performance crossover, for a smaller dataframe, it seems to be a win-win. But for larger, there is performance decrease in some cases.

Change	Before [4fe49b1]	After [93ea131] <fix_gh55509_v2~14>	Ratio	Benchmark (Parameter)
+	40.4±1ms	50.7±0.9ms	1.25	frame_methods.Update.time_to_update_big_frame_small_arg
-	68.1±0.8ms	35.9±0.9ms	0.53	frame_methods.Update.time_to_update_random_indices
-	26.4±0.2ms	5.98±0.07ms	0.23	frame_methods.Update.time_to_update_small_frame_big_arg

[rhshadrach]

lgtm

[rhshadrach]

Thanks @aureliobarbosa!