[Streaming] retry on requests errors

[lhoestq]

reported in https://discuss.huggingface.co/t/speeding-up-streaming-of-large-datasets-fineweb/90714/6 when training using a streaming a dataloader

cc @Wauplin it looks like the retries from hfh are not always enough. In this PR I let datasets do additional retries (that users can configure in datasets.config) since I couldn't find an easy way to increase the max_retries for hfh users in general.

[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.

[Wauplin]

(nit) Those two correspond to aiohttp.client_exceptions.ClientError and asyncio.TimeoutError. I think it's best to explicitly name them as you did with requests.exceptions.ConnectTimeout. It was not obvious for me at first (I wondered what was the difference between those 4 only to figure out it was not only requests errors)

[Wauplin]

Looks good yes 👍 Added a minor comment for clarity but that's all. huggingface_hub don't really have the concept of retrying on each and every HTTP request (only download requests) so I'm not surprised it can fail.

[lhoestq]

ci failures are r-unrelated to this PR, merging

[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)	0.005532 / 0.011353 (-0.005821)	0.004018 / 0.011008 (-0.006991)	0.064685 / 0.038508 (0.026177)	0.031303 / 0.023109 (0.008194)	0.254670 / 0.275898 (-0.021228)	0.271357 / 0.323480 (-0.052123)	0.003372 / 0.007986 (-0.004614)	0.004153 / 0.004328 (-0.000175)	0.050381 / 0.004250 (0.046131)	0.046837 / 0.037052 (0.009784)	0.253166 / 0.258489 (-0.005323)	0.294257 / 0.293841 (0.000416)	0.029746 / 0.128546 (-0.098800)	0.012519 / 0.075646 (-0.063127)	0.208822 / 0.419271 (-0.210449)	0.036925 / 0.043533 (-0.006608)	0.247636 / 0.255139 (-0.007503)	0.269102 / 0.283200 (-0.014097)	0.019021 / 0.141683 (-0.122662)	1.138825 / 1.452155 (-0.313330)	1.203301 / 1.492716 (-0.289415)

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)	0.095950 / 0.018006 (0.077944)	0.303347 / 0.000490 (0.302857)	0.000221 / 0.000200 (0.000022)	0.000042 / 0.000054 (-0.000012)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.019014 / 0.037411 (-0.018397)	0.062220 / 0.014526 (0.047694)	0.074811 / 0.176557 (-0.101745)	0.122917 / 0.737135 (-0.614218)	0.075765 / 0.296338 (-0.220574)

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)	0.288359 / 0.215209 (0.073150)	2.849491 / 2.077655 (0.771837)	1.479448 / 1.504120 (-0.024672)	1.350560 / 1.541195 (-0.190635)	1.366079 / 1.468490 (-0.102411)	0.733609 / 4.584777 (-3.851168)	2.416014 / 3.745712 (-1.329698)	2.954834 / 5.269862 (-2.315028)	1.985703 / 4.565676 (-2.579974)	0.080589 / 0.424275 (-0.343686)	0.005581 / 0.007607 (-0.002026)	0.343706 / 0.226044 (0.117661)	3.416257 / 2.268929 (1.147329)	1.865937 / 55.444624 (-53.578687)	1.545911 / 6.876477 (-5.330566)	1.711004 / 2.142072 (-0.431069)	0.821231 / 4.805227 (-3.983996)	0.138865 / 6.500664 (-6.361799)	0.046466 / 0.075469 (-0.029003)

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)	0.965632 / 1.841788 (-0.876155)	11.812101 / 8.074308 (3.737792)	9.399156 / 10.191392 (-0.792236)	0.143325 / 0.680424 (-0.537099)	0.014824 / 0.534201 (-0.519377)	0.306143 / 0.579283 (-0.273140)	0.264063 / 0.434364 (-0.170301)	0.347820 / 0.540337 (-0.192517)	0.476818 / 1.386936 (-0.910118)

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)	0.005978 / 0.011353 (-0.005375)	0.004482 / 0.011008 (-0.006526)	0.053788 / 0.038508 (0.015280)	0.033963 / 0.023109 (0.010853)	0.267258 / 0.275898 (-0.008640)	0.290916 / 0.323480 (-0.032563)	0.004485 / 0.007986 (-0.003500)	0.002876 / 0.004328 (-0.001453)	0.048637 / 0.004250 (0.044386)	0.042050 / 0.037052 (0.004997)	0.278607 / 0.258489 (0.020118)	0.315411 / 0.293841 (0.021570)	0.032059 / 0.128546 (-0.096487)	0.012851 / 0.075646 (-0.062795)	0.061672 / 0.419271 (-0.357600)	0.034545 / 0.043533 (-0.008988)	0.262068 / 0.255139 (0.006929)	0.291197 / 0.283200 (0.007997)	0.019092 / 0.141683 (-0.122591)	1.108690 / 1.452155 (-0.343464)	1.161025 / 1.492716 (-0.331691)

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)	0.096775 / 0.018006 (0.078768)	0.306825 / 0.000490 (0.306335)	0.000210 / 0.000200 (0.000010)	0.000054 / 0.000054 (-0.000000)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.023160 / 0.037411 (-0.014251)	0.078794 / 0.014526 (0.064268)	0.088954 / 0.176557 (-0.087602)	0.129488 / 0.737135 (-0.607648)	0.091239 / 0.296338 (-0.205099)

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)	0.292911 / 0.215209 (0.077702)	2.910802 / 2.077655 (0.833148)	1.569310 / 1.504120 (0.065191)	1.433807 / 1.541195 (-0.107388)	1.478619 / 1.468490 (0.010129)	0.720982 / 4.584777 (-3.863795)	0.972104 / 3.745712 (-2.773608)	3.026941 / 5.269862 (-2.242921)	1.919170 / 4.565676 (-2.646506)	0.079292 / 0.424275 (-0.344983)	0.005227 / 0.007607 (-0.002380)	0.345363 / 0.226044 (0.119319)	3.416149 / 2.268929 (1.147221)	1.938377 / 55.444624 (-53.506248)	1.626037 / 6.876477 (-5.250440)	1.644405 / 2.142072 (-0.497668)	0.802485 / 4.805227 (-4.002742)	0.135114 / 6.500664 (-6.365550)	0.042015 / 0.075469 (-0.033454)

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)	1.014812 / 1.841788 (-0.826976)	12.583844 / 8.074308 (4.509536)	10.522495 / 10.191392 (0.331103)	0.143336 / 0.680424 (-0.537088)	0.015843 / 0.534201 (-0.518357)	0.306556 / 0.579283 (-0.272727)	0.129654 / 0.434364 (-0.304710)	0.340442 / 0.540337 (-0.199896)	0.445220 / 1.386936 (-0.941716)