CARTESIAN (left outer) (anti) semi join may cause OOM even if right table is not very large

[gengliqi]

Bug Report

For CARTESIAN (left outer) (anti) semi join, each row in the left table's block must be combined with all of the rows in the right table's block in order to execute other_conditions expression or/and other_eq_conditions_from_in expression.

After that, according to the result of expressions, each row in the left table's block will only produce one or zero number of rows to the result block through filter function as below.

tiflash/dbms/src/Interpreters/Join.cpp

Line 1460 in 4206a52

block.getByPosition(i).column = block.getByPosition(i).column->filter(row_filter, -1);

tiflash/dbms/src/Interpreters/Join.cpp

Line 1545 in 4206a52

block.getByPosition(i).column = block.getByPosition(i).column->filter(row_filter, -1);

 
Then the issue comes. The allocated memory size of these columns may be much greater than the used memory size of them because filter will use the size of source column to reserve.

tiflash/dbms/src/Columns/IColumn.h

Lines 215 to 222 in 4206a52

	/** Removes elements that don't match the filter.
	* Is used in WHERE and HAVING operations.
	* If result_size_hint > 0, then makes advance reserve(result_size_hint) for the result column;
	* if 0, then don't makes reserve(),
	* otherwise (i.e. < 0), makes reserve() using size of source column.
	*/
	using Filter = PaddedPODArray<UInt8>;
	virtual Ptr filter(const Filter & filt, ssize_t result_size_hint) const = 0;

 
For example, suppose the number of rows in the right table is 50k, left_rows_per_iter will be only 1 when max_block_size_for_cross_join is 64k.

tiflash/dbms/src/Interpreters/Join.cpp

Lines 1890 to 1898 in 4206a52

	size_t left_rows_per_iter = std::max(rows_left, 1);
	if (max_block_size_for_cross_join > 0 && right_table_rows > 0 && other_condition_ptr != nullptr
	&& CrossJoinAdder<KIND, STRICTNESS>::allRightRowsMaybeAdded())
	{
	/// if other_condition is not null, and all right columns maybe added during join, try to use multiple iter
	/// to make memory usage under control, for anti semi cross join that is converted by not in subquery,
	/// it is likely that other condition may filter out most of the rows
	left_rows_per_iter = std::max(max_block_size_for_cross_join / right_table_rows, 1);
	}

So in each iteration, there should be only one or zero number of rows in the result block while the allocated memory size is approximately equal to the memory of 50k rows.

If there are 64k rows in this block of the left table, there are 64k result blocks and the allocated memory size is approximately equal to the memory of 50k * 64k = 3276800k rows, which is likely to make TiFlash OOM.

[LittleFall]

Please remove the may-affect-** tag as needed or convert it to a definitive affect-** tag to make pr pass the ci check