feat: support non streamable arrow file binary format

[kmehant]

Support Arrow files (.arrow) that are in non streamable binary file formats.

[kmehant]

requesting review - @albertvillanova @lhoestq

[lhoestq]

Awesome thank you ! this will be pretty useful :)

Before we merge could you also add a test in tests/packaged_modules/test_arrow.py ?

I noticed it's pretty empty right now compared to test_json.py or test_csv.py though, maybe I can take care of it next week if needed

[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.

[kmehant]

@lhoestq rebased the PR, It would be really helpful to have this feature into datasets, please let me know if there is anything pending on this PR, thanks.

[kmehant]

@lhoestq

Have added the unit test to generate tables for both the arrow formats - file and streaming.

Let me know if we have any docs changes as well. Thanks

[kmehant]

@lhoestq any update on this thread? Thanks

[prince14322]

Timely PR!
Can we please look into this?

[albertvillanova]

Thank you for the useful enhancement and the test!

[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)	0.005737 / 0.011353 (-0.005615)	0.003894 / 0.011008 (-0.007114)	0.067510 / 0.038508 (0.029002)	0.033431 / 0.023109 (0.010321)	0.262766 / 0.275898 (-0.013132)	0.283776 / 0.323480 (-0.039704)	0.003296 / 0.007986 (-0.004689)	0.003577 / 0.004328 (-0.000752)	0.052165 / 0.004250 (0.047915)	0.047815 / 0.037052 (0.010763)	0.263528 / 0.258489 (0.005039)	0.292980 / 0.293841 (-0.000861)	0.031535 / 0.128546 (-0.097011)	0.012966 / 0.075646 (-0.062680)	0.218827 / 0.419271 (-0.200444)	0.039181 / 0.043533 (-0.004352)	0.263768 / 0.255139 (0.008629)	0.288012 / 0.283200 (0.004813)	0.020562 / 0.141683 (-0.121121)	1.180547 / 1.452155 (-0.271608)	1.269283 / 1.492716 (-0.223433)

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)	0.098951 / 0.018006 (0.080944)	0.318922 / 0.000490 (0.318433)	0.000214 / 0.000200 (0.000014)	0.000044 / 0.000054 (-0.000010)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.021315 / 0.037411 (-0.016097)	0.067728 / 0.014526 (0.053202)	0.079428 / 0.176557 (-0.097129)	0.127472 / 0.737135 (-0.609663)	0.080455 / 0.296338 (-0.215883)

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)	0.308725 / 0.215209 (0.093516)	3.043555 / 2.077655 (0.965900)	1.587419 / 1.504120 (0.083299)	1.444421 / 1.541195 (-0.096774)	1.470703 / 1.468490 (0.002213)	0.784005 / 4.584777 (-3.800772)	2.582064 / 3.745712 (-1.163648)	3.140269 / 5.269862 (-2.129592)	2.031099 / 4.565676 (-2.534577)	0.086999 / 0.424275 (-0.337277)	0.005923 / 0.007607 (-0.001684)	0.361333 / 0.226044 (0.135289)	3.587173 / 2.268929 (1.318244)	1.961448 / 55.444624 (-53.483177)	1.649868 / 6.876477 (-5.226609)	1.698595 / 2.142072 (-0.443478)	0.858552 / 4.805227 (-3.946676)	0.146001 / 6.500664 (-6.354663)	0.046049 / 0.075469 (-0.029421)

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)	1.022644 / 1.841788 (-0.819144)	12.655994 / 8.074308 (4.581686)	10.205832 / 10.191392 (0.014440)	0.156073 / 0.680424 (-0.524351)	0.015550 / 0.534201 (-0.518651)	0.327762 / 0.579283 (-0.251521)	0.299212 / 0.434364 (-0.135152)	0.367549 / 0.540337 (-0.172788)	0.474499 / 1.386936 (-0.912437)

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)	0.005904 / 0.011353 (-0.005448)	0.004245 / 0.011008 (-0.006763)	0.054309 / 0.038508 (0.015801)	0.037490 / 0.023109 (0.014381)	0.293540 / 0.275898 (0.017642)	0.324068 / 0.323480 (0.000588)	0.004675 / 0.007986 (-0.003311)	0.003091 / 0.004328 (-0.001238)	0.052972 / 0.004250 (0.048721)	0.045545 / 0.037052 (0.008493)	0.301465 / 0.258489 (0.042976)	0.342822 / 0.293841 (0.048981)	0.033958 / 0.128546 (-0.094588)	0.013311 / 0.075646 (-0.062336)	0.064050 / 0.419271 (-0.355222)	0.038127 / 0.043533 (-0.005406)	0.297383 / 0.255139 (0.042244)	0.312244 / 0.283200 (0.029044)	0.019395 / 0.141683 (-0.122288)	1.244335 / 1.452155 (-0.207820)	1.305547 / 1.492716 (-0.187169)

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)	0.101847 / 0.018006 (0.083840)	0.330827 / 0.000490 (0.330337)	0.000211 / 0.000200 (0.000011)	0.000047 / 0.000054 (-0.000008)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.025734 / 0.037411 (-0.011677)	0.085020 / 0.014526 (0.070494)	0.096724 / 0.176557 (-0.079833)	0.141276 / 0.737135 (-0.595859)	0.099150 / 0.296338 (-0.197189)

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)	0.316058 / 0.215209 (0.100849)	3.059459 / 2.077655 (0.981804)	1.638394 / 1.504120 (0.134274)	1.505313 / 1.541195 (-0.035881)	1.526635 / 1.468490 (0.058145)	0.777259 / 4.584777 (-3.807518)	1.059575 / 3.745712 (-2.686137)	2.952334 / 5.269862 (-2.317528)	2.003894 / 4.565676 (-2.561782)	0.084464 / 0.424275 (-0.339811)	0.007343 / 0.007607 (-0.000265)	0.366218 / 0.226044 (0.140174)	3.705588 / 2.268929 (1.436660)	2.047029 / 55.444624 (-53.397595)	1.766970 / 6.876477 (-5.109507)	1.883804 / 2.142072 (-0.258268)	0.865780 / 4.805227 (-3.939447)	0.143180 / 6.500664 (-6.357485)	0.044943 / 0.075469 (-0.030527)

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)	1.141391 / 1.841788 (-0.700397)	13.244917 / 8.074308 (5.170609)	10.907863 / 10.191392 (0.716471)	0.156087 / 0.680424 (-0.524337)	0.016487 / 0.534201 (-0.517714)	0.331377 / 0.579283 (-0.247906)	0.148863 / 0.434364 (-0.285501)	0.370443 / 0.540337 (-0.169895)	0.499647 / 1.386936 (-0.887289)