[FLINK-33722] MATCH_RECOGNIZE in batch mode ignores events order

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 1.17.1
• Component/s: Library / CEP
• Labels: pull-request-available
• Assignee: Unassigned
• Reporter: Grzegorz Kołakowski
• Created: 01/Dec/23 09:57
• Updated: 22/Apr/24 14:52

Description

MATCH_RECOGNIZE in batch mode seems to ignore ORDER BY clause. Let's consider the following example:

    FROM events
        MATCH_RECOGNIZE (
            PARTITION BY user_id
            ORDER BY ts ASC
            MEASURES
                FIRST(A.ts) as _start,
                LAST(A.ts) as _middle,
                LAST(B.ts) as _finish
            ONE ROW PER MATCH
            AFTER MATCH SKIP PAST LAST ROW
            PATTERN (A{2} B) WITHIN INTERVAL '2' HOURS
            DEFINE
                A AS active is false,
                B AS active is true
        ) AS T 

where events is a Postgresql table containing ~10000 records.

CREATE TABLE events (
  id INT,
  user_id INT,
  ts TIMESTAMP(3),
  active BOOLEAN,
  WATERMARK FOR ts AS ts - INTERVAL '5' SECOND,
  PRIMARY KEY (id) NOT ENFORCED
) WITH (
    'connector' = 'jdbc',
    'url' = 'jdbc:postgresql://postgres:5432/test',
    'username' = 'test',
    'password' = 'test',
    'table-name' = 'events'
); 

It can happen that _finish is smaller than _start or _middle, which is wrong.

   user_id                  _start                 _middle                 _finish
         1 2023-11-23 14:34:42.346 2023-11-23 14:34:48.370 2023-11-23 14:34:44.264

 

Repository where I reproduced the problem: https://github.com/grzegorz8/flink-match-recognize-in-batch-debugging

---

 

According to dwysakowicz:  In BATCH the CepOperator is always created to process records in processing time:
https://github.com/apache/flink/blob/7f7bee70e3ac0d9fb27d7e09b41d6396b748dada/flink-table/flink-table-planner/src/main/java/org/apache/flink/table/planner/plan/nodes/exec/batch/BatchExecMatch.java#L54
A comparator is passed along to the operator covering the sorting on ts field: https://github.com/apache/flink/blob/fea9ffedecf81a97de5c31519ade3bab8228e743/flink-table/flink-table-planner/src/main/java/org/apache/flink/table/planner/plan/nodes/exec/common/CommonExecMatch.java#L173 but this is only secondary sorting. It is applied only within records of the same timestamp.

Issue Links

• relates to: FLINK-24865 Support MATCH_RECOGNIZE in Batch mode
• links to: GitHub Pull Request #24699