Implement Underlying Grouping Sets

[AilinKid]

Feature Request

Is your feature request related to a problem? Please describe:

Grouping Sets is internal implementation mechanism for supporting Multi-Distinct-Aggregate MPP Optimization and Rollup/Cube syntax.

SELECT a, b, sum(expression) FROM table GROUP BY a, b With Rollup;

For modern databases like Spark SQL, it allows user to explicitly describe wanted grouping sets explicitly like this:

SELECT a, b, sum(expression) FROM table GROUP BY a, b GROUPING SETS((a,b),(x,x),(...));

Different listed grouping sets/grouping layout requirement above will ask the underlying data to be expanded as multi copies to feed different requirement of Aggregation granularity. As a consequence, the leveled-aggregated result rows will be a union to user.

Apart from explicit requirement from sql syntax level, there is a another way to implicitly describe a composed grouping sets. That's what exactly rollup and cube syntax does. For more detail about a example like rollup(a,b,c), it has implicit N = 4 grouping sets derived from incremental expression composition, such as grouping sets (), (a),(a,b),(a,b,c), so does cube syntax which will be more complicated one.

For Multi Distinct Aggregate case like

select count(distinct a), count(distinct b) from t

distinct nature require a implement of aggregation on groups grouped by a or b here, while single one copy of data can't satisfied both grouping by a and grouping b synchronously. As a consequence, we resort to different grouping sets like (a) and (b) to ask the underlying data to be expanded to feed different aggregation vertically.

Both of the 3 cases above is dependent/based on the implementation of Grouping Sets and Expand Operator, so that's why this issue calls for.

Describe the feature you'd like:

Shown above.

Describe alternatives you've considered:

For Rollup Syntax workaround, rewrite the SQL as union of many sub-query with individual group by items.
For Multi Distinct Aggregate Optimization workaround, there is no way to migrate the computation task to multi mpp nodes.

Teachability, Documentation, Adoption, Migration Strategy:

related issues schedule

Underlying Grouping Sets and Expand Operator

• planner, expression: support multi-distinct agg under MPP mode #39973 multi distinct aggregate planner side based on Expand1
• executor: support the grouping sets tipb#283 grouping sets based Expand1 pb support
• plan, executor: implement Expand operator for grouping sets tiflash#6545 multi distinct aggregate tiflash side based on Expand1
• planner, executor: refactor expand logic from analyzing grouping sets to level projections tiflash#7169 leveled projections refactor of Expand2
• executor: add pb support for leveled projection of expand2 tipb#300 leveled projection Expand2 pb support

Rollup Syntax

• Add GroupingFunctionMetadata tipb#298 Grouping function pb support
• Prepare grouping function for roll up #42463 Grouping function TiDB side implementation
• Support the with rollup syntax for group clause #43427 Support Rollup syntax in TiDB parser side
• grouping function should support multi args like grouping(a,b,c)... tiflash#7590 grouping function should support multi args like grouping(a,b,c)

Infra Support

• expression: add simple expression semantic equal check logic #43558 Expression tree canonical semantic equal check
• Support basic expression(grouping sets) transformation for Rollup syntax #44112 Support Rollup grouping sets transformation
• expression: modify the grouping function to receive uint64 #44184 Modify the grouping function to receive uint64
• rollup expand should support being converted to pb.expand2  #45179 Physical expand support expand-pb protocol V2

Plan Operator

• planner: introduce the logical Expand for rollup syntax #44214 Introduce Logical Expand for rollup syntax
• support grouping function/col/expression rewriting and physical plan exhaustion for rollup expand OP #44487 Support grouping function/col/expression rewriting and physical plan exhaustion for rollup expand OP

Bug Fix

• function: fix grouping function result name tiflash#7718 function: fix grouping function result name
• having item ref-ed to a grouping expression shouldn't be push down through expand #45647 having item ref-ed to a grouping expression shouldn't be push down through expand
• order-by clause should be resolved to grouping set item #45593 order-by clause should be resolved to grouping set item
• rollup aggregation report tiflash column name mismatch #45576 rollup aggregation report tiflash column name mismatch
• expand: tpc-ds rollup related queries errors in nightly cluster tiflash#7830 expand: tpc-ds rollup related queries errors in nightly cluster
  • Illegal types String and Decimal(10,0) of arguments of function multiply when rollup #45276
  • Illegal column Nullable(Int32) of first argument of function tidb_cast when rollup #45275
  • Nullable(Int32) cannot be inside Nullable column when rollup #45273
  • Block schema mismatch in FineGrainedShuffleWriter-V1: different types: expected Nullable(String), got String when rollup #45272
  • rollup aggregation report tiflash column name mismatch #45576
• rollup expand should support being converted to pb.expand2  #45179 rollup expand should support being converted to pb.expand2
• grouping function arg check logic errors in pre-only-full-group-check phase #45661 function: defer grouping function validation logic to expression rewriter.
• orderby clause ref priority doesn't take effect #45797 orderby clause ref priority doesn't take effect
• grouping function can't just be recreated with function name and args because it has additional metadata #45756 grouping function can't just be recreated with function name and args because it has additional metadata