Child pages
  • Subqueries in SELECT
Skip to end of metadata
Go to start of metadata

Problem

Currently Hive doesn't support subqueries in a SELECT statement, for example, the following query will not run on Hive:

Recently a lot of work has been done to extend support for subqueries (HIVE-15456). But this work primarily targeted extending subquery support in WHERE and HAVING clauses. We plan to continue the work done in HIVE-15456 to support subqueries in a select list (see HIVE-16091).

Assumptions

We plan to limit the scope with the following assumptions and limitations.

  • Subqueries could only be top-level expressions in SELECT. That is, subqueries in complex expressions, aggregates, UDFs, etc. will not be supported for now. For example the following queries will not run on Hive:

    Not Supported
  • Scalar subqueries can only return at most one row. Hive will check for this case at runtime and throw an error if not satisfied. For example the following query is invalid:

    Not Supported
  • Scalar subqueries can only have one column. Hive will check for this case during compilation and throw an error. For example the following query is invalid:

    Not Supported
  • Correlated variables are only permitted in a filter, that is, a WHERE or HAVING clause. For example the following query is invalid:

    Not Supported
  • Subqueries with DISTINCT are not allowed. Since DISTINCT <expression> will be evaluated as GROUP BY <expression>, subqueries with DISTINCT are disallowed for now.

Design

Given the assumptions above, the following kind of subqueries could be used in SELECT. 

  • Scalar subqueries, for example: 

  • IN subqueries, for example:

  • EXISTS subqueries, for example:

All of the above queries could be correlated or uncorrelated.

Design for this will be similar to the work done in HIVE-15456.

  • genLogicalPlan will go over the select list to do the following:
    • If subquery is not a top-level expression, throw an error.
    • Otherwise, generate an appropriate plan by using RexSubquery to represent the subquery.
  • HiveSubqueryRemoveRule will then be applied to remove the RexSubquery node and rewrite the query into a join.
  • HiveRelDecorrelator::decorrelateQuery will then be used to decorrelate correlated queries.

 HIVE-16091 covers the initial work for supporting subqueries in SELECT.

  • No labels