Skip to content

Group-by without ordering on aggregate #11706

Open
@Jackie-Jiang

Description

When a group-by query does not have order-by on the aggregate column, we don't need to keep more groups than the LIMIT because the order-by value won't change. We can maintain a heap (PriorityQueue) of LIMIT values. On the group-key generation side, we should also keep only the relevant keys.

One common query is:
SELECT COUNT(*) FROM myTable GROUP BY timeCol ORDER BY timeCol DESC LIMIT 10

Problems to solve:

  1. Group-by query with order-by on the key column:
    Currently we keep Math.max(5000, LIMIT * 5) groups, which is not necessary since only the top LIMIT groups are relevant

  2. Group-by query without order-by:
    Currently we keep random LIMIT groups per server, and there is no guarantee the same group is picked across different servers, which can lead to wrong result when there are more than LIMIT groups

Solution:

  1. To ensure the ordering is deterministic (we need this guarantee to ensure the groups returned from all servers are the same), we should append all non-ordering group keys implicitly. This is one exception: when we want to keep all groups on the server, we don't need this since all groups will be returned anyway.
  2. Optimize the execution when all the ordering keys are group key

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Assignees

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions