Added fast path for validating ASCII text (~1.12-1.89x improvement on reading ASCII parquet data)

[Dandandan]

A bit inspired by a commit from @ritchie46 in polars.

We can use is_ascii for smaller strings, as it's much faster than the utf8 check, especially when applied on a full buffer.
For (larger) utf8 there is not a real perf hit.

Results on the benchmarks:

read utf8 2^10                 1.00     33.9±0.65µs    1.16     39.4±1.88µs
read utf8 2^12                 1.00     48.4±0.37µs    1.38     66.9±2.05µs
read utf8 2^14                 1.00    101.7±2.47µs    1.57    159.2±1.98µs
read utf8 2^16                 1.00    293.3±2.35µs    1.80   528.5±20.57µs
read utf8 2^18                 1.00   1056.6±9.58µs    1.89      2.0±0.02ms
read utf8 2^20                 1.00      4.4±0.02ms    1.81      8.0±0.12ms
read utf8 dict 2^10            1.00     36.7±1.12µs    1.14     41.9±1.10µs
read utf8 dict 2^12            1.00     55.9±1.40µs    1.27     71.2±1.87µs
read utf8 dict 2^14            1.00    123.2±0.77µs    1.51    186.4±2.38µs
read utf8 dict 2^16            1.00    360.8±3.16µs    1.74    627.9±7.52µs
read utf8 dict 2^18            1.00  1340.1±10.74µs    1.74      2.3±0.03ms
read utf8 dict 2^20            1.00      5.4±0.07ms    1.71      9.2±0.10ms
read utf8 large 2^10           1.00     54.0±1.15µs    1.10     59.3±1.44µs
read utf8 large 2^12           1.00    131.0±1.93µs    1.05    137.1±4.08µs
read utf8 large 2^14           1.00   911.1±19.27µs    1.05   959.5±14.10µs
read utf8 large 2^16           1.01      3.9±0.03ms    1.00      3.9±0.10ms
read utf8 large 2^18           1.00     17.1±0.16ms    1.06     18.1±0.40ms
read utf8 large 2^20           1.00     69.8±0.54ms    1.11     77.4±1.07ms
read utf8 multi 2^10           1.00     37.1±0.31µs    1.12     41.6±0.81µs
read utf8 multi 2^12           1.00     54.2±0.42µs    1.39     75.6±5.97µs
read utf8 multi 2^14           1.00    110.9±1.12µs    1.55    171.7±2.59µs
read utf8 multi 2^16           1.00   358.4±11.54µs    1.63    582.9±7.18µs
read utf8 multi 2^18           1.00   1258.4±7.18µs    1.83      2.3±0.04ms
read utf8 multi 2^20           1.00      5.1±0.05ms    1.68      8.6±0.10ms
read utf8 multi snappy 2^10    1.00     33.9±0.82µs    1.18     39.8±0.81µs
read utf8 multi snappy 2^12    1.00     52.3±0.44µs    1.30     68.0±1.97µs
read utf8 multi snappy 2^14    1.00    124.9±4.95µs    1.53    190.9±5.59µs
read utf8 multi snappy 2^16    1.00    385.8±7.92µs    1.68    646.7±6.01µs
read utf8 multi snappy 2^18    1.00   1429.2±6.74µs    1.73      2.5±0.22ms
read utf8 multi snappy 2^20    1.00      6.0±0.05ms    1.60      9.6±0.33ms

[Dandandan]

FYI @ritchie46 this implementation should be faster than writing a manual loop with all for larger buffers.

[codecov]

Codecov Report

Merging #542 (9ba4b0f) into main (002a0ef) will increase coverage by 0.02%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##             main     #542      +/-   ##
==========================================
+ Coverage   81.26%   81.28%   +0.02%     
==========================================
  Files         380      380              
  Lines       23382    23386       +4     
==========================================
+ Hits        19001    19010       +9     
+ Misses       4381     4376       -5     

Impacted Files	Coverage Δ
src/array/specification.rs	83.33% <100.00%> (+2.08%)	⬆️
src/io/ipc/write/common.rs	76.28% <0.00%> (+0.64%)	⬆️
src/bitmap/utils/slice_iterator.rs	94.02% <0.00%> (+1.49%)	⬆️
src/compute/arithmetics/time.rs	46.93% <0.00%> (+2.04%)	⬆️
src/io/avro/read/schema.rs	55.95% <0.00%> (+2.38%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 002a0ef...9ba4b0f. Read the comment docs.

[jorgecarleitao]

Looks great! Thanks a lot @Dandandan ! Could you rebase on top of the latest main? I merged #543 just to safeguard us against edge cases ^_^

[Dandandan]

Looks great! Thanks a lot @Dandandan ! Could you rebase on top of the latest main? I merged #543 just to safeguard us against edge cases ^_^

Done!

[jorgecarleitao]

benchmarks on https://github.com/DataEngineeringLabs/parquet-benchmark updated. Solid improvements 👍 Thanks a lot @ritchie46 and @Dandandan . 2.5x faster than pyarrow.

I also just noticed that pyarrow uses multi-threading by default, so I still want to make the comparison a bit more fair. :)

[Dandandan]

benchmarks on https://github.com/DataEngineeringLabs/parquet-benchmark updated. Solid improvements 👍 Thanks a lot @ritchie46 and @Dandandan . 2.5x faster than pyarrow.

I also just noticed that pyarrow uses multi-threading by default, so I still want to make the comparison a bit more fair. :)

Nice, thanks @ritchie46 for the great idea.

I should rerun some benchmarks with DataFusion/arrow2 soon :)