[CASSANDRA-17811] Fix CQL aggregation functions for collections, tuples and UDTs

• Type: New Feature
• Status: Resolved
• Priority: Normal
• Resolution: Fixed
• Fix Version/s: 5.0-alpha1, 5.0
• Component/s: CQL/Semantics
• Epic Link: CEP-20: Dynamic Data Masking
• Complexity: Normal
• Platform: All

Source Control Link

https://github.com/apache/cassandra/commit/6da9e33602fad4b8bf9466dc0e9a73665469a195

Test and Documentation Plan

New unit tests are included.

• Assignee: Andres de la Peña
• Reporter: Andres de la Peña

Authors

Andres de la Peña

Reviewers

Benjamin Lerer

• Created: 11/Aug/22 16:37
• Updated: 12/Sep/23 13:02
• Resolved: 18/Nov/22 10:39

Description

It has been found during CASSANDRA-8877 that CQLS's aggregation functions max, min and count can be applied to collections, but the result is returned as a blob. For example:

CREATE TABLE t (k int PRIMARY KEY, l list<int>);
INSERT INTO t(k, l) VALUES (0, [1, 2, 3]);
INSERT INTO t(k, l) VALUES (1, [10, 20, 30]);
SELECT max(l) FROM t;

 system.max(l)
------------------------------------------------------------
 0x00000003000000040000000a0000000400000014000000040000001e

This happens on 3.0, 3.11, 4.0, 4.1 and trunk.

I'm not sure on whether the function shouldn't be supported for collections, or it should be supported but the result is wrong.

In the example above, the returned blob is the serialized value of [10, 20, 30], which is the right one according to the list comparator. I think this happens because the matched version of the function is the one for (blob) -> blob. We would need a (list<int>) -> list<int> function instead, but this function doesn't exist.

It would be quite easy to add versions of the max, min and count functions for every type of collection (list<int>, list<text>, map<int, int>, map<int, text>, etc.). The downside of this approach is that it would increase the number of aggregation functions kept in memory from 82 to 2722, if my maths are right. This is quite an increase, mainly due to the many possible combinations of the map type. Here is a quick, incomplete prototype of the approach.

Also, I'm not sure that applying those aggregation functions to collections is very useful in practice. Thus, an alternative approach would be just forbidding them, considering them not supported. I don't think it would be a problem for backward compatibility since no one has complained about the current behaviour, and we might well consider that the original intent was not to allow aggregation on collections. At least, there aren't any tests for it, and I can't find any documentation about it either.

Another idea that comes to mind is that we could change the meaning of those functions to aggregate the values within the collection, instead of aggregating the rows. In that case, the behaviour would be:

CREATE TABLE t (k int PRIMARY KEY, l list<int>);
INSERT INTO t(k, l) VALUES (0, [1, 2, 3]);
INSERT INTO t(k, l) VALUES (1, [10, 20, 30]);
SELECT max(l) FROM t;

 k | system.max(l)
---+-----------
 1 | 30
 0 | 3

Of course we could have separate function names for that type of collection aggregations, like collectionMax, maxItem, or something like that.

Issue Links

• Discovered while testing: CASSANDRA-8877 Ability to read the TTL and WRITE TIME of an element in a collection
• is depended upon by: CASSANDRA-17941 CQL data masking functions
• Parent Feature: CASSANDRA-18060 Add aggregation scalar functions on collections
• links to: PR trunk