Use pandas ujson in JSON loader to improve performance

[albertvillanova]

Use pandas ujson in JSON loader to improve performance.

Note that datasets has pandas as required dependency. And pandas includes ujson in pd.io.json.ujson_loads.

Fix #6867.

CC: @natolambert

[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.

[albertvillanova]

Before pandas-2.2.0, the function ujson_loads was named loads: https://github.com/pandas-dev/pandas/blob/v2.1.0/pandas/io/json/__init__.py#L5

import ujson_loads as loads

[lhoestq]

Great ! do you know what the speed-up is (roughly) ?

[albertvillanova]

Thanks for your review, @lhoestq.

The performance gain depends on many factors, such as underlying data structures, file size...

In my benchmark, the performance gain was around 8.1%.

[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)	0.005428 / 0.011353 (-0.005925)	0.003682 / 0.011008 (-0.007326)	0.064360 / 0.038508 (0.025852)	0.032044 / 0.023109 (0.008934)	0.238281 / 0.275898 (-0.037617)	0.267542 / 0.323480 (-0.055937)	0.003152 / 0.007986 (-0.004834)	0.003292 / 0.004328 (-0.001037)	0.050157 / 0.004250 (0.045906)	0.048311 / 0.037052 (0.011259)	0.253743 / 0.258489 (-0.004746)	0.282729 / 0.293841 (-0.011112)	0.027271 / 0.128546 (-0.101275)	0.010238 / 0.075646 (-0.065408)	0.208179 / 0.419271 (-0.211092)	0.035607 / 0.043533 (-0.007925)	0.246750 / 0.255139 (-0.008389)	0.263362 / 0.283200 (-0.019837)	0.018475 / 0.141683 (-0.123208)	1.152978 / 1.452155 (-0.299177)	1.158545 / 1.492716 (-0.334171)

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)	0.096645 / 0.018006 (0.078639)	0.313186 / 0.000490 (0.312696)	0.000209 / 0.000200 (0.000009)	0.000052 / 0.000054 (-0.000003)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.018800 / 0.037411 (-0.018612)	0.065833 / 0.014526 (0.051307)	0.073668 / 0.176557 (-0.102888)	0.120608 / 0.737135 (-0.616527)	0.074936 / 0.296338 (-0.221403)

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)	0.281596 / 0.215209 (0.066387)	2.814537 / 2.077655 (0.736882)	1.482781 / 1.504120 (-0.021338)	1.349770 / 1.541195 (-0.191424)	1.371571 / 1.468490 (-0.096919)	0.555068 / 4.584777 (-4.029709)	2.369588 / 3.745712 (-1.376124)	2.742771 / 5.269862 (-2.527091)	1.711519 / 4.565676 (-2.854158)	0.060921 / 0.424275 (-0.363354)	0.005263 / 0.007607 (-0.002344)	0.333721 / 0.226044 (0.107677)	3.329598 / 2.268929 (1.060669)	1.806983 / 55.444624 (-53.637641)	1.515730 / 6.876477 (-5.360746)	1.557622 / 2.142072 (-0.584451)	0.619564 / 4.805227 (-4.185663)	0.115503 / 6.500664 (-6.385161)	0.041728 / 0.075469 (-0.033741)

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)	0.967300 / 1.841788 (-0.874487)	11.295081 / 8.074308 (3.220773)	9.535119 / 10.191392 (-0.656273)	0.140232 / 0.680424 (-0.540192)	0.013774 / 0.534201 (-0.520427)	0.281847 / 0.579283 (-0.297436)	0.260076 / 0.434364 (-0.174288)	0.323657 / 0.540337 (-0.216681)	0.421116 / 1.386936 (-0.965820)

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)	0.005276 / 0.011353 (-0.006077)	0.003639 / 0.011008 (-0.007370)	0.050451 / 0.038508 (0.011943)	0.032787 / 0.023109 (0.009678)	0.267029 / 0.275898 (-0.008869)	0.299899 / 0.323480 (-0.023581)	0.004177 / 0.007986 (-0.003809)	0.002697 / 0.004328 (-0.001631)	0.049631 / 0.004250 (0.045380)	0.041942 / 0.037052 (0.004889)	0.279249 / 0.258489 (0.020760)	0.306512 / 0.293841 (0.012671)	0.029340 / 0.128546 (-0.099207)	0.010118 / 0.075646 (-0.065528)	0.058243 / 0.419271 (-0.361028)	0.033871 / 0.043533 (-0.009662)	0.265949 / 0.255139 (0.010810)	0.284263 / 0.283200 (0.001064)	0.017351 / 0.141683 (-0.124332)	1.107081 / 1.452155 (-0.345074)	1.184946 / 1.492716 (-0.307770)

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)	0.095621 / 0.018006 (0.077614)	0.304758 / 0.000490 (0.304269)	0.000204 / 0.000200 (0.000004)	0.000052 / 0.000054 (-0.000003)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.022444 / 0.037411 (-0.014967)	0.075894 / 0.014526 (0.061368)	0.089077 / 0.176557 (-0.087480)	0.126960 / 0.737135 (-0.610176)	0.089120 / 0.296338 (-0.207218)

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)	0.289885 / 0.215209 (0.074676)	2.843219 / 2.077655 (0.765565)	1.582704 / 1.504120 (0.078584)	1.426551 / 1.541195 (-0.114644)	1.431591 / 1.468490 (-0.036899)	0.577265 / 4.584777 (-4.007512)	0.956040 / 3.745712 (-2.789673)	2.753517 / 5.269862 (-2.516345)	1.732503 / 4.565676 (-2.833173)	0.063511 / 0.424275 (-0.360764)	0.005089 / 0.007607 (-0.002518)	0.339205 / 0.226044 (0.113160)	3.339148 / 2.268929 (1.070219)	1.901543 / 55.444624 (-53.543081)	1.618392 / 6.876477 (-5.258084)	1.612885 / 2.142072 (-0.529188)	0.656563 / 4.805227 (-4.148664)	0.116740 / 6.500664 (-6.383924)	0.040497 / 0.075469 (-0.034973)

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)	1.005568 / 1.841788 (-0.836219)	11.872770 / 8.074308 (3.798462)	9.867118 / 10.191392 (-0.324274)	0.130193 / 0.680424 (-0.550231)	0.022857 / 0.534201 (-0.511344)	0.281908 / 0.579283 (-0.297375)	0.125978 / 0.434364 (-0.308386)	0.382604 / 0.540337 (-0.157733)	0.415078 / 1.386936 (-0.971858)