gpupreagg: Improve num_groups estimation for arrow_fdw with sqrt-base…

[a-hirota]

Problem

PostgreSQL multiplies n_distinct values for each GROUP BY column, causing overestimation. This makes GpuPreAgg less likely to be selected by the planner.

Revised: For arrow_fdw/parquet_fdw, running ANALYZE requires decompressing Parquet files. This change makes GpuPreAgg more likely to trigger even with default estimates (i.e., without ANALYZE).

Solution

• Apply sqrt dampening: adjusted = baseline + sqrt(excess) * factor
• Configurable via pg_strom.groupby_reduction_factor GUC (default 0.7, range 0.1-2.0)
• Baseline set to max(input_rows * 0.1, 1000)
• Only applies when selectivity > 10% and estimated_groups > 10K
• Only activated for arrow_fdw/parquet_fdw foreign tables

Example with factor=0.7:
estimated=1M, input=1M, baseline=100K adjusted = 100K + sqrt(900K) * 0.7 ≈ 101K

[Copilot]

Pull Request Overview

This PR improves GROUP BY cardinality estimation for arrow_fdw/parquet_fdw foreign tables by implementing square-root based dampening to reduce overestimated group counts that make GpuPreAgg less likely to be selected by the PostgreSQL planner.

• Implements sqrt-based reduction algorithm to address PostgreSQL's tendency to overestimate num_groups by multiplying n_distinct values across GROUP BY columns
• Adds configurable dampening via new GUC parameter pg_strom.groupby_reduction_factor (default 0.7, range 0.1-2.0)
• Restricts the optimization to arrow_fdw/parquet_fdw foreign tables only to avoid requiring expensive ANALYZE operations

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[Copilot]

Potential division by zero if input_nrows is 0. Add a guard to check if input_nrows > 0 before performing the division.

Suggested change

	/* Guard against division by zero */
	if (input_nrows <= 0.0)
	return estimated_groups;

[Copilot]

Corrected spacing in format string - removed extra space before opening parenthesis.

Suggested change

	elog(DEBUG1, "GpuPreAgg: reduced num_groups from %.0f to %.0f "
	elog(DEBUG1, "GpuPreAgg: reduced num_groups from %.0f to %.0f"

[Copilot]

[nitpick] This block-level comment should be moved above the opening brace of the block it describes (line 2324) for better code organization and readability.

Suggested change

	/* Adjust num_groups if it seems overestimated due to combinatorial explosion.
	* Apply square-root based dampening to reduce overestimated group counts.
	* Only applies to arrow_fdw/parquet_fdw foreign tables (because ANALYZE
	* requires decompression which consumes GPU memory, so we avoid ANALYZE
	* and fix the default cardinality estimation instead).
	*/
	{
	{
	/* Adjust num_groups if it seems overestimated due to combinatorial explosion.
	* Apply square-root based dampening to reduce overestimated group counts.
	* Only applies to arrow_fdw/parquet_fdw foreign tables (because ANALYZE
	* requires decompression which consumes GPU memory, so we avoid ANALYZE
	* and fix the default cardinality estimation instead).
	*/

[a-hirota]

● Copilotから3つの指摘を対応しました。履歴をクリーンに保ちたいのでforce-pushしました。

1. Division by Zero防止: input_nrows > 0 のチェックを追加
2. フォーマット修正: debug logの末尾の余分なスペースを削除
3. コメント配置: ブロックコメントを開き括弧の上に移動