POC: Window function intra-operator parallelism and vectorization (up to 50× faster)

[2010YOUY01]

Which issue does this PR close?

POC for #22355

Rationale for this change

This optimization only targets a specific type of window functions:

• Fixed ROWS window frame
• Regular aggregate function

See Issue for the implementation idea:

• Optimize window function with intra-operator parallelism and vectorization #22355

Benchmark

In datafusion-cli:

create external table lineitem
stored as parquet
location '/Users/yongting/Code/datafusion/benchmarks/data/tpch_sf10/lineitem'
with order (l_orderkey asc);

-- PR only config
set datafusion.optimizer.prefer_window_agg_exec=true;

-- Q1: Single global partition, no sort, low scanning overhead
SELECT
  v1,
  avg(v1) OVER (
    ORDER BY v1
    ROWS BETWEEN 3 PRECEDING AND 3 FOLLOWING
  ) AS moving_avg
FROM generate_series(1000000000) AS t1(v1);

-- Q2: Single global partition, window expr matches parquet existing order
SELECT
  l_orderkey,
  avg(l_partkey) OVER (
    ORDER BY l_orderkey
    ROWS BETWEEN 1000 PRECEDING AND 1000 FOLLOWING
  ) AS moving_avg
FROM lineitem;

-- Q3: 3 partitions
SELECT
  l_orderkey,
  avg(l_partkey) OVER (
    PARTITION BY l_returnflag
    ORDER BY l_orderkey
    ROWS BETWEEN 100 PRECEDING AND CURRENT ROW
  ) AS moving_avg
FROM lineitem;

-- Q4: 100k partitions
SELECT
  l_orderkey,
  avg(l_partkey) OVER (
    PARTITION BY l_suppkey
    ORDER BY l_orderkey
    ROWS BETWEEN 3 PRECEDING AND 3 FOLLOWING
  ) AS moving_avg
FROM lineitem;

Result

| Query | PR   | Main | Speedup |
|-------|------|------|----------|
| Q1    | 1.40 | 79   | 56.43x  |
| Q2    | 0.99 | 5.44 | 5.49x   |
| Q3    | 2.66 | 5.15 | 1.94x   |
| Q4    | 1.08 | 1.55 | 1.44x   |

(Q3 and Q4 include ~1s of sorting time, so the window execution is likely much faster)

What changes are included in this PR?

Are these changes tested?

Are there any user-facing changes?

[2010YOUY01]

This PoC only targets a specific window expression case. I am still prototyping and thinking through a design that is easy to extend to other possible optimizations for different window function types. After I figure that out, I will split the work into smaller PRs.