GPORCA Limitations

There are limitations in Apache Cloudberry when using the default GPORCA optimizer. GPORCA and the Postgres-based planner currently coexist in Apache Cloudberry because GPORCA does not support all Apache Cloudberry features.

This section describes the limitations.

• Unsupported SQL Query Features
• Performance Regressions

Unsupported SQL query features

Certain query features are not supported with the default GPORCA optimizer. When an unsupported query is run, Cloudberry logs this notice along with the query text:

Falling back to Postgres-based planner because GPORCA does not support the following feature: UTILITY command

These features are unsupported when GPORCA is enabled (the default):

• SP-GiST indexing method. GPORCA supports only B-tree, bitmap, GIN, and GiST indexes. GPORCA ignores indexes created with unsupported methods.

• The SELECT command's TABLESAMPLE clause.

• The optional WITH ORDINALITY clause of a function call inside the FROM clause of a SELECT command.

• Multi-level partitioned tables.

• Non-uniform partitioned tables.

• SortMergeJoin (SMJ).

• Ordered aggregations.

• Multi-argument DISTINCT qualified aggregates, for example SELECT corr(DISTINCT a, b) FROM tbl1;.

• Multiple grouping sets specified using a duplicate alias in a null-producing grouping set spec. Such queries fall back to the Postgres-based planner unless you directly coerce the alias to a separate variable as shown in the example below:

  CREATE TEMP TABLE tempt AS SELECT i AS ai1, i AS ai2 FROM generate_series(1, 3)i;
  SELECT ai1, ai2 FROM tempt GROUP BY ai2, ROLLUP(ai1) ORDER BY ai1, ai2;

• These scalar operators:

  • ROW
  • ROWCOMPARE
  • FIELDSELECT

• Aggregate functions that take set operators as input arguments.

• Multiple Distinct Qualified Aggregates, such as SELECT count(DISTINCT a), sum(DISTINCT b) FROM foo, are not supported by default. They can be enabled with the optimizer_enable_multiple_distinct_aggs Configuration Parameter.

• percentile_* window functions (ordered-set aggregate functions).

• Inverse distribution functions.

• Queries that run functions that are defined with the ON COORDINATOR or ON ALL SEGMENTS attribute.

• Queries that contain UNICODE characters in metadata names, such as table names, and the characters are not compatible with the host system locale.

• SELECT, UPDATE, and DELETE commands where a table name is qualified by the ONLY keyword.

• Per-column collation. GPORCA supports collation only when all columns in the query use the same collation. If columns in the query use different collations, then Cloudberry uses the Postgres-based planner.

• DML and COPY ... FROM operations on foreign tables.

Performance regressions

The following features are known performance regressions that occur with GPORCA enabled:

• Short running queries - For GPORCA, short running queries might encounter additional overhead due to GPORCA enhancements for determining an optimal query execution plan.
• ANALYZE - For GPORCA, the ANALYZE command generates root partition statistics for partitioned tables. For the Postgres-based planner, these statistics are not generated.
• DML operations - For GPORCA, DML enhancements including the support of updates on partition and distribution keys might require additional overhead.

Also, enhanced functionality of the features from previous versions could result in additional time required when GPORCA runs SQL statements with the features.