Fast table queries with interpolation search

[lhoestq]

Intro

This should fix issue #1803

Currently querying examples in a dataset is O(n) because of the underlying pyarrow ChunkedArrays implementation.
To fix this I implemented interpolation search that is pretty effective since datasets usually verifies the condition of evenly distributed chunks (the default chunk size is fixed).

Benchmark

Here is a benchmark I did on bookcorpus (74M rows):

for the current implementation

>>> python speed.py
Loaded dataset 'bookcorpus', len=74004228, nbytes=4835358766


========================= Querying unshuffled bookcorpus =========================

Avg access time key=1                                                 : 0.018ms
Avg access time key=74004227                                          : 0.215ms
Avg access time key=range(74003204, 74004228)                         : 1.416ms
Avg access time key=RandIter(low=0, high=74004228, size=1024, seed=42): 92.532ms

========================== Querying shuffled bookcorpus ==========================

Avg access time key=1                                                 : 0.187ms
Avg access time key=74004227                                          : 6.642ms
Avg access time key=range(74003204, 74004228)                         : 90.941ms
Avg access time key=RandIter(low=0, high=74004228, size=1024, seed=42): 3448.456ms

for the new one using interpolation search:

>>> python speed.py
Loaded dataset 'bookcorpus', len=74004228, nbytes=4835358766


========================= Querying unshuffled bookcorpus =========================

Avg access time key=1                                                 : 0.076ms
Avg access time key=74004227                                          : 0.056ms
Avg access time key=range(74003204, 74004228)                         : 1.807ms
Avg access time key=RandIter(low=0, high=74004228, size=1024, seed=42): 24.028ms

========================== Querying shuffled bookcorpus ==========================

Avg access time key=1                                                 : 0.061ms
Avg access time key=74004227                                          : 0.058ms
Avg access time key=range(74003204, 74004228)                         : 22.166ms
Avg access time key=RandIter(low=0, high=74004228, size=1024, seed=42): 42.757ms

The RandIter class is just an iterable of 1024 random indices from 0 to 74004228.

Here is also a plot showing the speed improvement depending on the dataset size:

Implementation details:

• datasets.table.Table objects implement interpolation search for the slice method
• The interpolation search requires to store the offsets of all the chunks of a table. The offsets are stored when the Table is initialized.
• datasets.table.Table.slice returns a datasets.table.Table using interpolation search
• datasets.table.Table.fast_slice returns a pyarrow.Table object using interpolation search. This is useful to get a part of a dataset if we don't need the indexing structure for future computations. For example it's used when querying an example as a dictionary.
• Now a Dataset object is always backed by a datasets.table.Table object. If one passes a pyarrow.Table to initialize a Dataset, then it's converted to a datasets.table.Table

Checklist:

• implement interpolation search
• use datasets.table.Table in Dataset objects
• update current tests
• add tests for interpolation search
• comments and docstring
• add the benchmark to the CI

Fix #1803.