PyArrow 13 CI fixes

[mariosasko]

Fixes:

• bumps the PyArrow version check in the cast_array_to_feature to avoid the offset bug (still not fixed)
• aligns the Pandas formatting tests with the Numpy ones (the current test fails due to GH-33321: [Python] Support converting to non-nano datetime64 for pandas >= 2.0 apache/arrow#35656, which requires .to_pandas(coerce_temporal_nanoseconds=True) to always return datetime [ns] objects)

Fix #6173

[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)	0.006095 / 0.011353 (-0.005258)	0.003580 / 0.011008 (-0.007429)	0.080146 / 0.038508 (0.041638)	0.063445 / 0.023109 (0.040336)	0.321930 / 0.275898 (0.046032)	0.397933 / 0.323480 (0.074453)	0.003455 / 0.007986 (-0.004531)	0.002856 / 0.004328 (-0.001472)	0.062938 / 0.004250 (0.058687)	0.048896 / 0.037052 (0.011843)	0.333070 / 0.258489 (0.074581)	0.404485 / 0.293841 (0.110644)	0.027156 / 0.128546 (-0.101390)	0.007974 / 0.075646 (-0.067672)	0.261505 / 0.419271 (-0.157766)	0.045328 / 0.043533 (0.001795)	0.311203 / 0.255139 (0.056064)	0.390006 / 0.283200 (0.106806)	0.023650 / 0.141683 (-0.118033)	1.468856 / 1.452155 (0.016701)	1.503867 / 1.492716 (0.011151)

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)	0.202110 / 0.018006 (0.184103)	0.436433 / 0.000490 (0.435944)	0.002278 / 0.000200 (0.002078)	0.000070 / 0.000054 (0.000016)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.024575 / 0.037411 (-0.012836)	0.073005 / 0.014526 (0.058479)	0.083609 / 0.176557 (-0.092947)	0.144881 / 0.737135 (-0.592254)	0.083495 / 0.296338 (-0.212844)

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)	0.398911 / 0.215209 (0.183702)	3.994035 / 2.077655 (1.916381)	2.056768 / 1.504120 (0.552649)	1.913242 / 1.541195 (0.372047)	1.932934 / 1.468490 (0.464444)	0.498953 / 4.584777 (-4.085824)	3.031107 / 3.745712 (-0.714605)	2.817165 / 5.269862 (-2.452696)	1.858886 / 4.565676 (-2.706790)	0.056977 / 0.424275 (-0.367299)	0.006634 / 0.007607 (-0.000973)	0.472580 / 0.226044 (0.246536)	4.738301 / 2.268929 (2.469372)	2.373938 / 55.444624 (-53.070686)	2.021057 / 6.876477 (-4.855420)	2.195419 / 2.142072 (0.053346)	0.585182 / 4.805227 (-4.220045)	0.124260 / 6.500664 (-6.376405)	0.060250 / 0.075469 (-0.015219)

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)	1.227350 / 1.841788 (-0.614438)	18.496525 / 8.074308 (10.422216)	13.946658 / 10.191392 (3.755266)	0.140024 / 0.680424 (-0.540399)	0.017077 / 0.534201 (-0.517124)	0.334415 / 0.579283 (-0.244868)	0.351118 / 0.434364 (-0.083246)	0.379556 / 0.540337 (-0.160782)	0.525064 / 1.386936 (-0.861872)

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)	0.006176 / 0.011353 (-0.005177)	0.003648 / 0.011008 (-0.007360)	0.063461 / 0.038508 (0.024953)	0.062770 / 0.023109 (0.039660)	0.448786 / 0.275898 (0.172888)	0.486490 / 0.323480 (0.163010)	0.005527 / 0.007986 (-0.002458)	0.002860 / 0.004328 (-0.001469)	0.063803 / 0.004250 (0.059553)	0.049657 / 0.037052 (0.012604)	0.449625 / 0.258489 (0.191136)	0.489378 / 0.293841 (0.195537)	0.028406 / 0.128546 (-0.100140)	0.008062 / 0.075646 (-0.067584)	0.068417 / 0.419271 (-0.350854)	0.040854 / 0.043533 (-0.002678)	0.461670 / 0.255139 (0.206531)	0.481622 / 0.283200 (0.198423)	0.021018 / 0.141683 (-0.120665)	1.450328 / 1.452155 (-0.001826)	1.501283 / 1.492716 (0.008567)

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)	0.269824 / 0.018006 (0.251817)	0.412296 / 0.000490 (0.411807)	0.039582 / 0.000200 (0.039382)	0.000266 / 0.000054 (0.000211)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.026436 / 0.037411 (-0.010976)	0.080633 / 0.014526 (0.066107)	0.089786 / 0.176557 (-0.086770)	0.145020 / 0.737135 (-0.592115)	0.092327 / 0.296338 (-0.204012)

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)	0.464349 / 0.215209 (0.249140)	4.630631 / 2.077655 (2.552976)	2.560527 / 1.504120 (1.056407)	2.374195 / 1.541195 (0.833000)	2.424774 / 1.468490 (0.956284)	0.510428 / 4.584777 (-4.074349)	3.099805 / 3.745712 (-0.645907)	2.781096 / 5.269862 (-2.488765)	1.854276 / 4.565676 (-2.711400)	0.058102 / 0.424275 (-0.366173)	0.006365 / 0.007607 (-0.001242)	0.534082 / 0.226044 (0.308038)	5.355003 / 2.268929 (3.086074)	3.012546 / 55.444624 (-52.432078)	2.665222 / 6.876477 (-4.211255)	2.821014 / 2.142072 (0.678942)	0.597733 / 4.805227 (-4.207494)	0.125433 / 6.500664 (-6.375231)	0.060802 / 0.075469 (-0.014667)

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)	1.345699 / 1.841788 (-0.496088)	18.836083 / 8.074308 (10.761774)	14.895458 / 10.191392 (4.704066)	0.146843 / 0.680424 (-0.533581)	0.018082 / 0.534201 (-0.516119)	0.335729 / 0.579283 (-0.243554)	0.351013 / 0.434364 (-0.083351)	0.388435 / 0.540337 (-0.151902)	0.543826 / 1.386936 (-0.843110)

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[lhoestq]

Cool ! Just one question:

[lhoestq]

Should we have something smarter ? like a function that tests if the pyarrow bug is still there and caches the result

[mariosasko]

IMO this bug is too specific to run such a check.

[lhoestq]

It would be quite simple to check though no ? And it can be run only once the first time we need to check if the bug exists.

Anyway let's merge this PR ^^

[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)	0.006593 / 0.011353 (-0.004760)	0.004089 / 0.011008 (-0.006919)	0.084753 / 0.038508 (0.046245)	0.079899 / 0.023109 (0.056790)	0.311528 / 0.275898 (0.035630)	0.349722 / 0.323480 (0.026243)	0.004288 / 0.007986 (-0.003698)	0.004552 / 0.004328 (0.000224)	0.065896 / 0.004250 (0.061646)	0.053813 / 0.037052 (0.016760)	0.316958 / 0.258489 (0.058469)	0.367011 / 0.293841 (0.073170)	0.031082 / 0.128546 (-0.097464)	0.008684 / 0.075646 (-0.066963)	0.288003 / 0.419271 (-0.131268)	0.052560 / 0.043533 (0.009027)	0.305589 / 0.255139 (0.050450)	0.349656 / 0.283200 (0.066457)	0.023857 / 0.141683 (-0.117826)	1.462360 / 1.452155 (0.010205)	1.568170 / 1.492716 (0.075454)

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)	0.272342 / 0.018006 (0.254336)	0.585108 / 0.000490 (0.584618)	0.003427 / 0.000200 (0.003227)	0.000078 / 0.000054 (0.000023)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.030347 / 0.037411 (-0.007064)	0.086325 / 0.014526 (0.071799)	0.100958 / 0.176557 (-0.075598)	0.156534 / 0.737135 (-0.580601)	0.102506 / 0.296338 (-0.193832)

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)	0.406625 / 0.215209 (0.191416)	4.065957 / 2.077655 (1.988302)	2.075867 / 1.504120 (0.571747)	1.914390 / 1.541195 (0.373196)	2.013321 / 1.468490 (0.544831)	0.486832 / 4.584777 (-4.097945)	3.545940 / 3.745712 (-0.199772)	3.323226 / 5.269862 (-1.946635)	2.067742 / 4.565676 (-2.497934)	0.057884 / 0.424275 (-0.366391)	0.007751 / 0.007607 (0.000144)	0.484923 / 0.226044 (0.258878)	4.844885 / 2.268929 (2.575956)	2.569828 / 55.444624 (-52.874796)	2.224058 / 6.876477 (-4.652419)	2.485587 / 2.142072 (0.343515)	0.584311 / 4.805227 (-4.220916)	0.134984 / 6.500664 (-6.365680)	0.062164 / 0.075469 (-0.013305)

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)	1.247182 / 1.841788 (-0.594605)	20.107500 / 8.074308 (12.033192)	14.194444 / 10.191392 (4.003052)	0.147134 / 0.680424 (-0.533290)	0.018062 / 0.534201 (-0.516138)	0.392029 / 0.579283 (-0.187254)	0.402991 / 0.434364 (-0.031373)	0.457600 / 0.540337 (-0.082737)	0.632553 / 1.386936 (-0.754383)

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)	0.006920 / 0.011353 (-0.004433)	0.004257 / 0.011008 (-0.006751)	0.065233 / 0.038508 (0.026725)	0.078151 / 0.023109 (0.055042)	0.389141 / 0.275898 (0.113243)	0.431518 / 0.323480 (0.108038)	0.005752 / 0.007986 (-0.002234)	0.003584 / 0.004328 (-0.000745)	0.065173 / 0.004250 (0.060922)	0.059113 / 0.037052 (0.022060)	0.398225 / 0.258489 (0.139736)	0.430980 / 0.293841 (0.137139)	0.032802 / 0.128546 (-0.095744)	0.008702 / 0.075646 (-0.066945)	0.071345 / 0.419271 (-0.347926)	0.048269 / 0.043533 (0.004736)	0.389264 / 0.255139 (0.134125)	0.416008 / 0.283200 (0.132809)	0.024845 / 0.141683 (-0.116838)	1.499100 / 1.452155 (0.046945)	1.576397 / 1.492716 (0.083681)

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)	0.296674 / 0.018006 (0.278668)	0.540108 / 0.000490 (0.539619)	0.004293 / 0.000200 (0.004093)	0.000151 / 0.000054 (0.000096)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.034108 / 0.037411 (-0.003303)	0.092747 / 0.014526 (0.078221)	0.112203 / 0.176557 (-0.064354)	0.162728 / 0.737135 (-0.574407)	0.109955 / 0.296338 (-0.186383)

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)	0.432006 / 0.215209 (0.216797)	4.297591 / 2.077655 (2.219937)	2.379645 / 1.504120 (0.875525)	2.218680 / 1.541195 (0.677485)	2.314608 / 1.468490 (0.846117)	0.495562 / 4.584777 (-4.089215)	3.589787 / 3.745712 (-0.155925)	3.349593 / 5.269862 (-1.920268)	2.119893 / 4.565676 (-2.445783)	0.057976 / 0.424275 (-0.366299)	0.007612 / 0.007607 (0.000005)	0.509422 / 0.226044 (0.283378)	5.101444 / 2.268929 (2.832515)	2.794532 / 55.444624 (-52.650092)	2.459033 / 6.876477 (-4.417444)	2.714424 / 2.142072 (0.572352)	0.588444 / 4.805227 (-4.216784)	0.135763 / 6.500664 (-6.364901)	0.062593 / 0.075469 (-0.012876)

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)	1.361415 / 1.841788 (-0.480372)	20.940684 / 8.074308 (12.866376)	15.161364 / 10.191392 (4.969972)	0.154243 / 0.680424 (-0.526181)	0.020305 / 0.534201 (-0.513896)	0.397438 / 0.579283 (-0.181845)	0.415047 / 0.434364 (-0.019317)	0.473250 / 0.540337 (-0.067088)	0.740681 / 1.386936 (-0.646255)