Support streaming datasets with pyarrow.parquet.read_table #6251

albertvillanova · 2023-09-20T08:07:02Z

Support streaming datasets with pyarrow.parquet.read_table.

See: https://huggingface.co/datasets/uonlp/CulturaX/discussions/2

HuggingFaceDocBuilderDev · 2023-09-20T08:13:58Z

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

mariosasko · 2023-09-20T19:30:00Z

This function reads an entire Arrow table in one go, which is not ideal memory-wise, so I don't think we should encourage using this function, considering we want to keep RAM usage as low as possible in the streaming mode.

(Note that Parquet files are compressed, meaning the loaded table can be significantly larger than the size in Parquet.)

Instead, we should suggest the authors to use:

with open(doc_path, "rb") as f:
    parquet_file = pq.ParquetFile(f)
    for batch in parquet_file.iter_batches():
        pa_table = pa.Table.from_batches([batch])
        yield idx, pa_table
        idx += 1

albertvillanova · 2023-09-21T08:02:55Z

@mariosasko I think the potential problem you evoke is independent of whether or not we support streaming mode:

if the user's script with read_table works in non-streaming mode, it will also work in streaming mode after this PR

In fact, what we should suggest instead is to follow the scriptless approach, so that our parquet packaged module is used, with all the optimizations implemented. But this approach is not possible in all cases, and some use cases need to implement a script. And if they have small Parquet files and use read_table, I think we should support streaming.

In summary, let me clarify the goal and the scope of this PR:

a user needs using a loading script
their files are small enough so that they use read_table
their loading script works in non-streaming mode
therefore, this PR allows loading their dataset in streaming mode as well

mariosasko · 2023-09-21T14:19:04Z

Yes, the no-script approach with metadata configs makes the most sense.

their files are small enough so that they use read_table

Some of the Parquet files in that repo are larger than 1 GB ...

Also, I'd wait for more instances of people using the read_table function on the Hub before merging this PR.

albertvillanova · 2023-09-21T14:42:02Z

@mariosasko, yes, this solution is not specifically for the "uonlp/CulturaX" dataset, but for other use cases as I explained above: indeed, they finally removed the use of read_table because their data files are too large.

Also, I'd wait for more instances of people using the read_table function on the Hub before merging this PR.

Do you know how many datasets are currently using read_table?

mariosasko · 2023-09-21T17:12:52Z

Do you know how many datasets are currently using read_table?

Zero (based on the script that checks the script contents of the public Hub datasets).

albertvillanova · 2023-09-22T06:00:02Z

I see... Thanks! 🤗

albertvillanova · 2023-09-22T15:32:13Z

@mariosasko thanks for pointing the script! 🤗

However, I have found some Hub datasets that are using read_table, e.g.:

albertvillanova · 2023-09-27T06:26:17Z

I'm merging this PR as discussed in private.

github-actions · 2023-09-27T06:37:03Z

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)	0.008267 / 0.011353 (-0.003086)	0.005813 / 0.011008 (-0.005195)	0.108802 / 0.038508 (0.070294)	0.093996 / 0.023109 (0.070886)	0.403115 / 0.275898 (0.127217)	0.457299 / 0.323480 (0.133819)	0.006277 / 0.007986 (-0.001709)	0.004701 / 0.004328 (0.000373)	0.080700 / 0.004250 (0.076449)	0.077906 / 0.037052 (0.040854)	0.409972 / 0.258489 (0.151483)	0.477707 / 0.293841 (0.183867)	0.041816 / 0.128546 (-0.086731)	0.011250 / 0.075646 (-0.064397)	0.390634 / 0.419271 (-0.028637)	0.065361 / 0.043533 (0.021828)	0.404501 / 0.255139 (0.149362)	0.448162 / 0.283200 (0.164962)	0.032823 / 0.141683 (-0.108860)	1.899892 / 1.452155 (0.447737)	2.044561 / 1.492716 (0.551844)

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)	0.241093 / 0.018006 (0.223086)	0.482111 / 0.000490 (0.481622)	0.005505 / 0.000200 (0.005305)	0.000094 / 0.000054 (0.000039)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.034861 / 0.037411 (-0.002551)	0.109296 / 0.014526 (0.094770)	0.127594 / 0.176557 (-0.048962)	0.191815 / 0.737135 (-0.545320)	0.122630 / 0.296338 (-0.173709)

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)	0.452194 / 0.215209 (0.236985)	4.486200 / 2.077655 (2.408545)	2.155635 / 1.504120 (0.651515)	2.004569 / 1.541195 (0.463374)	2.142570 / 1.468490 (0.674080)	0.561488 / 4.584777 (-4.023289)	4.381102 / 3.745712 (0.635390)	3.914920 / 5.269862 (-1.354942)	2.474271 / 4.565676 (-2.091406)	0.067528 / 0.424275 (-0.356747)	0.008723 / 0.007607 (0.001116)	0.536077 / 0.226044 (0.310033)	5.342050 / 2.268929 (3.073122)	2.735747 / 55.444624 (-52.708877)	2.353938 / 6.876477 (-4.522539)	2.442878 / 2.142072 (0.300805)	0.685404 / 4.805227 (-4.119823)	0.156657 / 6.500664 (-6.344007)	0.071714 / 0.075469 (-0.003755)

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)	1.562852 / 1.841788 (-0.278935)	24.538203 / 8.074308 (16.463895)	16.857777 / 10.191392 (6.666385)	0.184221 / 0.680424 (-0.496203)	0.021688 / 0.534201 (-0.512513)	0.470700 / 0.579283 (-0.108583)	0.470593 / 0.434364 (0.036229)	0.645066 / 0.540337 (0.104729)	0.756075 / 1.386936 (-0.630861)

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)	0.009486 / 0.011353 (-0.001867)	0.004694 / 0.011008 (-0.006314)	0.080216 / 0.038508 (0.041708)	0.093479 / 0.023109 (0.070369)	0.537353 / 0.275898 (0.261455)	0.551631 / 0.323480 (0.228151)	0.007373 / 0.007986 (-0.000613)	0.004044 / 0.004328 (-0.000285)	0.075301 / 0.004250 (0.071051)	0.069408 / 0.037052 (0.032355)	0.527962 / 0.258489 (0.269473)	0.559423 / 0.293841 (0.265582)	0.039351 / 0.128546 (-0.089195)	0.010801 / 0.075646 (-0.064845)	0.092803 / 0.419271 (-0.326468)	0.058876 / 0.043533 (0.015343)	0.513742 / 0.255139 (0.258603)	0.574666 / 0.283200 (0.291466)	0.030277 / 0.141683 (-0.111406)	1.884936 / 1.452155 (0.432782)	2.008260 / 1.492716 (0.515543)

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)	0.242162 / 0.018006 (0.224156)	0.467400 / 0.000490 (0.466910)	0.005348 / 0.000200 (0.005148)	0.000103 / 0.000054 (0.000049)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.038022 / 0.037411 (0.000611)	0.108239 / 0.014526 (0.093713)	0.121514 / 0.176557 (-0.055042)	0.184951 / 0.737135 (-0.552184)	0.123138 / 0.296338 (-0.173200)

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)	0.558587 / 0.215209 (0.343377)	5.740312 / 2.077655 (3.662657)	3.172164 / 1.504120 (1.668044)	2.852908 / 1.541195 (1.311713)	2.894435 / 1.468490 (1.425945)	0.586399 / 4.584777 (-3.998378)	4.498342 / 3.745712 (0.752630)	4.000569 / 5.269862 (-1.269292)	2.610988 / 4.565676 (-1.954688)	0.068415 / 0.424275 (-0.355860)	0.008602 / 0.007607 (0.000994)	0.614731 / 0.226044 (0.388686)	6.068158 / 2.268929 (3.799229)	3.301070 / 55.444624 (-52.143554)	2.868034 / 6.876477 (-4.008443)	2.959072 / 2.142072 (0.816999)	0.684174 / 4.805227 (-4.121053)	0.154099 / 6.500664 (-6.346565)	0.070641 / 0.075469 (-0.004828)

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)	1.835667 / 1.841788 (-0.006120)	24.981645 / 8.074308 (16.907337)	17.218517 / 10.191392 (7.027125)	0.197055 / 0.680424 (-0.483368)	0.025465 / 0.534201 (-0.508736)	0.523498 / 0.579283 (-0.055785)	0.528268 / 0.434364 (0.093904)	0.599518 / 0.540337 (0.059180)	0.887206 / 1.386936 (-0.499730)

Support streaming pyarrow.parquet.read_table

08c75e0

Do not patch pyarrow attribute in packaged modules

2d95121

albertvillanova merged commit dd786d3 into huggingface:main Sep 27, 2023
10 of 12 checks passed

albertvillanova deleted the stream-pq-read-table branch September 27, 2023 06:26

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Support streaming datasets with pyarrow.parquet.read_table #6251

Support streaming datasets with pyarrow.parquet.read_table #6251

albertvillanova commented Sep 20, 2023

HuggingFaceDocBuilderDev commented Sep 20, 2023 •

edited

Loading

mariosasko commented Sep 20, 2023

albertvillanova commented Sep 21, 2023

mariosasko commented Sep 21, 2023

albertvillanova commented Sep 21, 2023

mariosasko commented Sep 21, 2023

albertvillanova commented Sep 22, 2023

albertvillanova commented Sep 22, 2023 •

edited

Loading

albertvillanova commented Sep 27, 2023

github-actions bot commented Sep 27, 2023

Benchmark: benchmark_array_xd.json

Benchmark: benchmark_getitem_100B.json

Benchmark: benchmark_indices_mapping.json

Benchmark: benchmark_iterating.json

Benchmark: benchmark_map_filter.json

Benchmark: benchmark_array_xd.json

Benchmark: benchmark_getitem_100B.json

Benchmark: benchmark_indices_mapping.json

Benchmark: benchmark_iterating.json

Benchmark: benchmark_map_filter.json

Support streaming datasets with pyarrow.parquet.read_table #6251

Support streaming datasets with pyarrow.parquet.read_table #6251

Conversation

albertvillanova commented Sep 20, 2023

HuggingFaceDocBuilderDev commented Sep 20, 2023 • edited Loading

mariosasko commented Sep 20, 2023

albertvillanova commented Sep 21, 2023

mariosasko commented Sep 21, 2023

albertvillanova commented Sep 21, 2023

mariosasko commented Sep 21, 2023

albertvillanova commented Sep 22, 2023

albertvillanova commented Sep 22, 2023 • edited Loading

albertvillanova commented Sep 27, 2023

github-actions bot commented Sep 27, 2023

Benchmark: benchmark_array_xd.json

Benchmark: benchmark_getitem_100B.json

Benchmark: benchmark_indices_mapping.json

Benchmark: benchmark_iterating.json

Benchmark: benchmark_map_filter.json

Benchmark: benchmark_array_xd.json

Benchmark: benchmark_getitem_100B.json

Benchmark: benchmark_indices_mapping.json

Benchmark: benchmark_iterating.json

Benchmark: benchmark_map_filter.json

HuggingFaceDocBuilderDev commented Sep 20, 2023 •

edited

Loading

albertvillanova commented Sep 22, 2023 •

edited

Loading