perf: Batch parquet primitive decoding

[coastalwhite]

This is rather large change that changes quite fundamentally how HybridRLE and Parquet decoding works. There are now two important concepts that speed up the performance of the Parquet reader while utilizing less memory than before. This does how increase the complexity of the code.

First, for a benchmark using the NYC Yellow-Taxi dataset (we decode the whole dataset 100x). Here, we see the following results.

No maximum threads:

Benchmark 1: After Optimization
  Time (mean ± σ):      4.918 s ±  0.076 s    [User: 28.748 s, System: 2.486 s]
  Range (min … max):    4.819 s …  5.064 s    10 runs

Benchmark 2: Before Optimization
  Time (mean ± σ):      7.333 s ±  2.144 s    [User: 60.374 s, System: 3.054 s]
  Range (min … max):    5.416 s … 11.132 s    10 runs

Summary
  After Optimization ran
    1.49 ± 0.44 times faster than Before Optimization

Maximum threads = 1:

Benchmark 1: After Optimization
  Time (mean ± σ):     18.452 s ±  0.054 s    [User: 16.058 s, System: 2.325 s]
  Range (min … max):   18.332 s … 18.511 s    10 runs

Benchmark 2: Before Optimization
  Time (mean ± σ):     27.027 s ±  0.062 s    [User: 24.668 s, System: 2.271 s]
  Range (min … max):   26.912 s … 27.105 s    10 runs

Summary
  After Optimization ran
    1.46 ± 0.01 times faster than Before Optimization

This PR introduces the concepts of a Translator and a BatchedCollector.

The Translator trait maps from hybrid RLE encoded values to an arbitrary set of values. The HybridRLEDecoder will can then collect and call the translator with batches of values. This way we minimize the amount of iterator polls and we do not need to allocate any more on the heap except the output buffer. This does however mean that the whole HybridRLEDecoder needs to be aware of the Translator trait.

Furthermore, the HybridRLEDecoder can now itself buffer, instead of using the BufferedHybridRleDecoderIter that was used before. Again, this allows minimal memory consumption and prevents constant polling.

The BatchedCollector is essentially a wrapper for the Pushable trait that automatically optimizes sequential pushes of valid and invalid values. It also allows for efficient skipping of values.

Overall, this change significantly speeds up the Parquet reader and extensive testing was done to ensure that no invalid data gets produced. But it is difficult to test all edge-cases.

From here, we can start incorporating the BatchedCollector and Translator traits in more places. In general, the HybridRleDecoder iterator implementation should effectively never be used.

[codecov]

Codecov Report

Attention: Patch coverage is 59.76530% with 480 lines in your changes missing coverage. Please review.

Project coverage is 80.48%. Comparing base (27ac6cc) to head (2370b0f).
Report is 3 commits behind head on main.

Files	Patch %	Lines
...rs-parquet/src/parquet/encoding/hybrid_rle/fuzz.rs	35.68%	182 Missing ⚠️
...ars-parquet/src/parquet/encoding/hybrid_rle/mod.rs	54.31%	106 Missing ⚠️
...arquet/src/parquet/encoding/hybrid_rle/buffered.rs	57.60%	78 Missing ⚠️
crates/polars-arrow/src/bitmap/utils/iterator.rs	65.13%	38 Missing ⚠️
...polars-parquet/src/arrow/read/deserialize/utils.rs	78.33%	26 Missing ⚠️
...quet/src/parquet/encoding/hybrid_rle/translator.rs	86.42%	19 Missing ⚠️
...s-parquet/src/parquet/encoding/bitpacked/decode.rs	81.66%	11 Missing ⚠️
...parquet/src/arrow/read/deserialize/binary/basic.rs	0.00%	7 Missing ⚠️
...arquet/src/arrow/read/deserialize/binview/basic.rs	42.85%	4 Missing ⚠️
.../arrow/read/deserialize/fixed_size_binary/basic.rs	33.33%	4 Missing ⚠️
... and 2 more

Additional details and impacted files

@@            Coverage Diff             @@
##             main   #17462      +/-   ##
==========================================
- Coverage   80.58%   80.48%   -0.10%     
==========================================
  Files        1480     1483       +3     
  Lines      193682   194830    +1148     
  Branches     2765     2770       +5     
==========================================
+ Hits       156071   156810     +739     
- Misses      37103    37511     +408     
- Partials      508      509       +1     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.