Base parquet batch_size on parquet row group size

[lhoestq]

This allows to stream datasets like Major-TOM/Core-S2L2A which have row groups with few rows (one row is ~10MB). Previously the cold start would take a lot of time and OOM because it would download many row groups before yielding the first example.

I tried on OpenOrca and imagenet-hard and it does't affect overall throughput.

Even if the overall throughput doesn't change for datasets like imagenet-hard with big rows, note that it does create shorter and more frequent pauses to download the next row group. Though I find it fine because previously the pauses were less frequent but very long (downloading multiple row groups at a time)

[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]

Thanks for making this more flexible.

Apart from the clarification on the new behavior, this looks good to me.

[albertvillanova]

This check was not present before. Is this behavior change in purpose? What if not parquet_file.metadata.num_row_groups > 0?

[lhoestq]

Yes it's because now I check for the metadata of the first row group, so I must first check that there is at least one ^^

[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)	0.005490 / 0.011353 (-0.005863)	0.003709 / 0.011008 (-0.007299)	0.064192 / 0.038508 (0.025684)	0.029581 / 0.023109 (0.006472)	0.251086 / 0.275898 (-0.024812)	0.267306 / 0.323480 (-0.056174)	0.003074 / 0.007986 (-0.004912)	0.003340 / 0.004328 (-0.000988)	0.048820 / 0.004250 (0.044569)	0.045370 / 0.037052 (0.008318)	0.260384 / 0.258489 (0.001895)	0.284558 / 0.293841 (-0.009283)	0.027732 / 0.128546 (-0.100814)	0.010661 / 0.075646 (-0.064986)	0.213403 / 0.419271 (-0.205868)	0.036283 / 0.043533 (-0.007250)	0.250107 / 0.255139 (-0.005032)	0.265220 / 0.283200 (-0.017980)	0.021021 / 0.141683 (-0.120661)	1.112058 / 1.452155 (-0.340096)	1.169039 / 1.492716 (-0.323678)

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)	0.095008 / 0.018006 (0.077002)	0.303509 / 0.000490 (0.303019)	0.000233 / 0.000200 (0.000033)	0.000052 / 0.000054 (-0.000002)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.018224 / 0.037411 (-0.019187)	0.061366 / 0.014526 (0.046841)	0.073584 / 0.176557 (-0.102972)	0.119869 / 0.737135 (-0.617266)	0.074228 / 0.296338 (-0.222111)

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)	0.288147 / 0.215209 (0.072938)	2.824419 / 2.077655 (0.746764)	1.478530 / 1.504120 (-0.025590)	1.350127 / 1.541195 (-0.191067)	1.349622 / 1.468490 (-0.118868)	0.568058 / 4.584777 (-4.016719)	2.377494 / 3.745712 (-1.368218)	2.720767 / 5.269862 (-2.549094)	1.710763 / 4.565676 (-2.854914)	0.061498 / 0.424275 (-0.362778)	0.004893 / 0.007607 (-0.002715)	0.335633 / 0.226044 (0.109588)	3.380646 / 2.268929 (1.111717)	1.802436 / 55.444624 (-53.642188)	1.562737 / 6.876477 (-5.313739)	1.566267 / 2.142072 (-0.575806)	0.629058 / 4.805227 (-4.176169)	0.116307 / 6.500664 (-6.384357)	0.042174 / 0.075469 (-0.033295)

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)	0.950945 / 1.841788 (-0.890842)	11.279009 / 8.074308 (3.204701)	9.433251 / 10.191392 (-0.758141)	0.138964 / 0.680424 (-0.541460)	0.014155 / 0.534201 (-0.520046)	0.284065 / 0.579283 (-0.295218)	0.263301 / 0.434364 (-0.171063)	0.331932 / 0.540337 (-0.208406)	0.441656 / 1.386936 (-0.945280)

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)	0.005132 / 0.011353 (-0.006221)	0.003484 / 0.011008 (-0.007524)	0.049040 / 0.038508 (0.010532)	0.030254 / 0.023109 (0.007145)	0.277141 / 0.275898 (0.001243)	0.295242 / 0.323480 (-0.028238)	0.004295 / 0.007986 (-0.003690)	0.002632 / 0.004328 (-0.001696)	0.048540 / 0.004250 (0.044290)	0.044787 / 0.037052 (0.007734)	0.287736 / 0.258489 (0.029247)	0.313146 / 0.293841 (0.019305)	0.029340 / 0.128546 (-0.099206)	0.010204 / 0.075646 (-0.065442)	0.059058 / 0.419271 (-0.360214)	0.051033 / 0.043533 (0.007500)	0.274086 / 0.255139 (0.018947)	0.293048 / 0.283200 (0.009848)	0.019573 / 0.141683 (-0.122110)	1.174032 / 1.452155 (-0.278123)	1.227107 / 1.492716 (-0.265609)

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)	0.094896 / 0.018006 (0.076890)	0.303519 / 0.000490 (0.303029)	0.000223 / 0.000200 (0.000023)	0.000049 / 0.000054 (-0.000005)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.021495 / 0.037411 (-0.015917)	0.074234 / 0.014526 (0.059708)	0.086212 / 0.176557 (-0.090345)	0.125052 / 0.737135 (-0.612084)	0.087464 / 0.296338 (-0.208874)

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)	0.297098 / 0.215209 (0.081889)	2.970944 / 2.077655 (0.893289)	1.650101 / 1.504120 (0.145981)	1.532694 / 1.541195 (-0.008501)	1.513652 / 1.468490 (0.045162)	0.559614 / 4.584777 (-4.025163)	2.404848 / 3.745712 (-1.340865)	2.627851 / 5.269862 (-2.642011)	1.707550 / 4.565676 (-2.858127)	0.061821 / 0.424275 (-0.362454)	0.005012 / 0.007607 (-0.002595)	0.342462 / 0.226044 (0.116417)	3.401703 / 2.268929 (1.132774)	1.991632 / 55.444624 (-53.452993)	1.737706 / 6.876477 (-5.138771)	1.837457 / 2.142072 (-0.304616)	0.638845 / 4.805227 (-4.166383)	0.114773 / 6.500664 (-6.385891)	0.040175 / 0.075469 (-0.035294)

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)	1.038286 / 1.841788 (-0.803501)	11.885757 / 8.074308 (3.811448)	10.061530 / 10.191392 (-0.129862)	0.140824 / 0.680424 (-0.539600)	0.015080 / 0.534201 (-0.519121)	0.287992 / 0.579283 (-0.291291)	0.273498 / 0.434364 (-0.160866)	0.326478 / 0.540337 (-0.213860)	0.426900 / 1.386936 (-0.960036)