Revert the changes in arrow_writer.py from #6636

[bryant1410]

#6636 broke write_examples_on_file and write_batch from the class ArrowWriter. I'm undoing these changes. See #6663.

Note the current implementation doesn't keep the order of the columns and the schema, thus setting a wrong schema for each column.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[mariosasko]

Hi! We can't revert this as the "reverted" implementation has quadratic time complexity. Instead, let's fix it:

[bryant1410]

Hi! We can't revert this as the "reverted" implementation has quadratic time complexity. Instead, let's fix it:

I agree, but it's the implementation we have had so far. Why don't we:

1. Release a hotfix ASAP (since would be doing a revert, we know it works as before) so people can continue using this library fine since AFAIU right now mostly writing examples for people is broken.
2. Then, focus on still applying the performance improvement and release again

[mariosasko]

The fix is straightforward, so one patch release (after this PR is merged) is enough.

Btw, let's also add a test to tests/test_arrow_writer.py to avoid this issue in the future.

[bryant1410]

Btw, let's also add a test to tests/test_arrow_writer.py to avoid this issue in the future.

Would you mind adding such test, as you're more familiar with the codebase?

[bryant1410]

I feel like we should really avoid this extra copy, especially if the inner iterable is large.

[mariosasko]

This negligible optimization (write_examples_on_file and write_batch are called every writer_batch_size) goes against good code practices.

We wouldn't use Python for this project if we wanted to optimize every aspect of the API.

[bryant1410]

I'm not sure it's negligible. #6636's OP stated:

I work with bioinformatics data and often these tables have thousands and even tens of thousands of features.

We'd create a list of tens of thousands of strings for every batch, for every processing step (e.g., a map).

And it's easy to remove (just cols = batch_samples, instead of copying it into a list).

Among other things, this library is about large data processing efficiency, so I think it'd be nice to consider it.

[mariosasko]

Thanks!

(The CI failure is not related to the changes)

[github-actions]

Show benchmarks

PyArrow==8.0.0

Show updated benchmarks!

Benchmark: benchmark_array_xd.json

metric	read_batch_formatted_as_numpy after write_array2d	read_batch_formatted_as_numpy after write_flattened_sequence	read_batch_formatted_as_numpy after write_nested_sequence	read_batch_unformated after write_array2d	read_batch_unformated after write_flattened_sequence	read_batch_unformated after write_nested_sequence	read_col_formatted_as_numpy after write_array2d	read_col_formatted_as_numpy after write_flattened_sequence	read_col_formatted_as_numpy after write_nested_sequence	read_col_unformated after write_array2d	read_col_unformated after write_flattened_sequence	read_col_unformated after write_nested_sequence	read_formatted_as_numpy after write_array2d	read_formatted_as_numpy after write_flattened_sequence	read_formatted_as_numpy after write_nested_sequence	read_unformated after write_array2d	read_unformated after write_flattened_sequence	read_unformated after write_nested_sequence	write_array2d	write_flattened_sequence	write_nested_sequence
new / old (diff)	0.005083 / 0.011353 (-0.006270)	0.003697 / 0.011008 (-0.007311)	0.063302 / 0.038508 (0.024794)	0.028866 / 0.023109 (0.005757)	0.249987 / 0.275898 (-0.025911)	0.270803 / 0.323480 (-0.052677)	0.004096 / 0.007986 (-0.003890)	0.002752 / 0.004328 (-0.001577)	0.049156 / 0.004250 (0.044906)	0.042936 / 0.037052 (0.005884)	0.266907 / 0.258489 (0.008418)	0.291462 / 0.293841 (-0.002379)	0.027703 / 0.128546 (-0.100844)	0.011006 / 0.075646 (-0.064641)	0.206238 / 0.419271 (-0.213033)	0.035446 / 0.043533 (-0.008087)	0.248923 / 0.255139 (-0.006216)	0.264141 / 0.283200 (-0.019058)	0.017545 / 0.141683 (-0.124138)	1.157145 / 1.452155 (-0.295009)	1.199007 / 1.492716 (-0.293710)

Benchmark: benchmark_getitem_100B.json

metric	get_batch_of_1024_random_rows	get_batch_of_1024_rows	get_first_row	get_last_row
new / old (diff)	0.092741 / 0.018006 (0.074734)	0.299057 / 0.000490 (0.298567)	0.000211 / 0.000200 (0.000011)	0.000049 / 0.000054 (-0.000006)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.017936 / 0.037411 (-0.019475)	0.061552 / 0.014526 (0.047026)	0.072938 / 0.176557 (-0.103618)	0.118192 / 0.737135 (-0.618944)	0.074589 / 0.296338 (-0.221750)

Benchmark: benchmark_iterating.json

metric	read 5000	read 50000	read_batch 50000 10	read_batch 50000 100	read_batch 50000 1000	read_formatted numpy 5000	read_formatted pandas 5000	read_formatted tensorflow 5000	read_formatted torch 5000	read_formatted_batch numpy 5000 10	read_formatted_batch numpy 5000 1000	shuffled read 5000	shuffled read 50000	shuffled read_batch 50000 10	shuffled read_batch 50000 100	shuffled read_batch 50000 1000	shuffled read_formatted numpy 5000	shuffled read_formatted_batch numpy 5000 10	shuffled read_formatted_batch numpy 5000 1000
new / old (diff)	0.287186 / 0.215209 (0.071977)	2.795694 / 2.077655 (0.718039)	1.474386 / 1.504120 (-0.029734)	1.359065 / 1.541195 (-0.182130)	1.375295 / 1.468490 (-0.093196)	0.569448 / 4.584777 (-4.015329)	2.374428 / 3.745712 (-1.371284)	2.770198 / 5.269862 (-2.499663)	1.716346 / 4.565676 (-2.849330)	0.063173 / 0.424275 (-0.361102)	0.005031 / 0.007607 (-0.002576)	0.333197 / 0.226044 (0.107153)	3.271739 / 2.268929 (1.002811)	1.826406 / 55.444624 (-53.618218)	1.554537 / 6.876477 (-5.321939)	1.565927 / 2.142072 (-0.576146)	0.649796 / 4.805227 (-4.155431)	0.118371 / 6.500664 (-6.382293)	0.042536 / 0.075469 (-0.032933)

Benchmark: benchmark_map_filter.json

metric	filter	map fast-tokenizer batched	map identity	map identity batched	map no-op batched	map no-op batched numpy	map no-op batched pandas	map no-op batched pytorch	map no-op batched tensorflow
new / old (diff)	0.969882 / 1.841788 (-0.871906)	11.638201 / 8.074308 (3.563893)	9.759370 / 10.191392 (-0.432022)	0.128069 / 0.680424 (-0.552355)	0.013493 / 0.534201 (-0.520708)	0.287324 / 0.579283 (-0.291959)	0.267542 / 0.434364 (-0.166821)	0.320072 / 0.540337 (-0.220265)	0.421132 / 1.386936 (-0.965804)

PyArrow==latest

Show updated benchmarks!

Benchmark: benchmark_array_xd.json

metric	read_batch_formatted_as_numpy after write_array2d	read_batch_formatted_as_numpy after write_flattened_sequence	read_batch_formatted_as_numpy after write_nested_sequence	read_batch_unformated after write_array2d	read_batch_unformated after write_flattened_sequence	read_batch_unformated after write_nested_sequence	read_col_formatted_as_numpy after write_array2d	read_col_formatted_as_numpy after write_flattened_sequence	read_col_formatted_as_numpy after write_nested_sequence	read_col_unformated after write_array2d	read_col_unformated after write_flattened_sequence	read_col_unformated after write_nested_sequence	read_formatted_as_numpy after write_array2d	read_formatted_as_numpy after write_flattened_sequence	read_formatted_as_numpy after write_nested_sequence	read_unformated after write_array2d	read_unformated after write_flattened_sequence	read_unformated after write_nested_sequence	write_array2d	write_flattened_sequence	write_nested_sequence
new / old (diff)	0.005679 / 0.011353 (-0.005674)	0.003746 / 0.011008 (-0.007262)	0.050149 / 0.038508 (0.011641)	0.034382 / 0.023109 (0.011273)	0.289802 / 0.275898 (0.013904)	0.314993 / 0.323480 (-0.008487)	0.004488 / 0.007986 (-0.003498)	0.002786 / 0.004328 (-0.001542)	0.047987 / 0.004250 (0.043737)	0.046589 / 0.037052 (0.009537)	0.301420 / 0.258489 (0.042931)	0.335384 / 0.293841 (0.041543)	0.050701 / 0.128546 (-0.077845)	0.010987 / 0.075646 (-0.064660)	0.058292 / 0.419271 (-0.360979)	0.033973 / 0.043533 (-0.009560)	0.288923 / 0.255139 (0.033784)	0.306263 / 0.283200 (0.023064)	0.018856 / 0.141683 (-0.122827)	1.160721 / 1.452155 (-0.291433)	1.208151 / 1.492716 (-0.284565)

Benchmark: benchmark_getitem_100B.json

metric	get_batch_of_1024_random_rows	get_batch_of_1024_rows	get_first_row	get_last_row
new / old (diff)	0.092633 / 0.018006 (0.074626)	0.300353 / 0.000490 (0.299864)	0.000219 / 0.000200 (0.000019)	0.000045 / 0.000054 (-0.000009)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.022257 / 0.037411 (-0.015154)	0.075417 / 0.014526 (0.060892)	0.087289 / 0.176557 (-0.089268)	0.125416 / 0.737135 (-0.611720)	0.088751 / 0.296338 (-0.207588)

Benchmark: benchmark_iterating.json

metric	read 5000	read 50000	read_batch 50000 10	read_batch 50000 100	read_batch 50000 1000	read_formatted numpy 5000	read_formatted pandas 5000	read_formatted tensorflow 5000	read_formatted torch 5000	read_formatted_batch numpy 5000 10	read_formatted_batch numpy 5000 1000	shuffled read 5000	shuffled read 50000	shuffled read_batch 50000 10	shuffled read_batch 50000 100	shuffled read_batch 50000 1000	shuffled read_formatted numpy 5000	shuffled read_formatted_batch numpy 5000 10	shuffled read_formatted_batch numpy 5000 1000
new / old (diff)	0.286477 / 0.215209 (0.071268)	2.801931 / 2.077655 (0.724277)	1.553034 / 1.504120 (0.048914)	1.426152 / 1.541195 (-0.115043)	1.443824 / 1.468490 (-0.024666)	0.563298 / 4.584777 (-4.021479)	2.428968 / 3.745712 (-1.316744)	2.685964 / 5.269862 (-2.583897)	1.752304 / 4.565676 (-2.813372)	0.064174 / 0.424275 (-0.360101)	0.005079 / 0.007607 (-0.002528)	0.344899 / 0.226044 (0.118855)	3.372528 / 2.268929 (1.103600)	1.900723 / 55.444624 (-53.543901)	1.623721 / 6.876477 (-5.252756)	1.781009 / 2.142072 (-0.361064)	0.655229 / 4.805227 (-4.149998)	0.116050 / 6.500664 (-6.384614)	0.040374 / 0.075469 (-0.035095)

Benchmark: benchmark_map_filter.json

metric	filter	map fast-tokenizer batched	map identity	map identity batched	map no-op batched	map no-op batched numpy	map no-op batched pandas	map no-op batched pytorch	map no-op batched tensorflow
new / old (diff)	1.004714 / 1.841788 (-0.837074)	12.108179 / 8.074308 (4.033871)	10.233447 / 10.191392 (0.042055)	0.141438 / 0.680424 (-0.538986)	0.015387 / 0.534201 (-0.518814)	0.288068 / 0.579283 (-0.291216)	0.277025 / 0.434364 (-0.157339)	0.331714 / 0.540337 (-0.208623)	0.424209 / 1.386936 (-0.962727)