[CASSANDRA-19404] Unexpected NullPointerException in ANN+WHERE when adding rows in another partition

• Type: Bug
• Status: Resolved
• Priority: Normal
• Resolution: Fixed
• Fix Version/s: 5.0-beta2, 5.0, 5.1-alpha1, 5.1-alpha, 5.1
• Component/s: Feature/Vector Search
• Bug Category: Degradation - Other Exception
• Severity: Normal
• Complexity: Normal
• Discovered By: User Report
• Platform: All
• Since Version: 5.0-alpha1

Source Control Link

https://github.com/apache/cassandra/commit/54cf65476a83d8cec04486e787eaa82489871a4a. https://github.com/apache/cassandra/commit/cd347bdc5810654ea04cc45e0affa9304d0f1798

Test and Documentation Plan

5.0 tiny patch with the added test here - https://github.com/apache/cassandra/pull/3130

Green CI here (one container timeout, but I won't rerun it as we have enough repeated runs for the tests already) - https://app.circleci.com/pipelines/github/ekaterinadimitrova2/cassandra?branch=19404-5.0

• Assignee: Ekaterina Dimitrova
• Reporter: Stefano Lottini

Authors

Ekaterina Dimitrova

Reviewers

Andres de la Peña

• Created: 15/Feb/24 14:24
• Updated: 26/Mar/24 10:38
• Resolved: 29/Feb/24 14:40

Description

• Bug observed on the Docker image 5.0-beta1
• Bug also observed on latest head of Cassandra repo (as of 2024-02-15)
• (working fine on vsearch branch of datastax/cassandra, commit hash 80c2f8b9ad5b89efee0645977a5ca53943717c0d)

Summary: A query with ann + where clause on a map + where clause on the partition key starts erroring once there are other partitions in the table.

There are three SELECT statements in the repro minimal code below - the third is where the error is triggered.

// reproduced with Dockerized Cassandra 5.0-beta1 on 2024-02-15

/////////
// SCHEMA
/////////

CREATE TABLE ks.v_table (
    pk int,
    row_v vector<float, 2>,
    metadata map<text, text>,
    PRIMARY KEY (pk)
);
CREATE CUSTOM INDEX v_md
    ON ks.v_table (entries(metadata))
    USING 'StorageAttachedIndex';
CREATE CUSTOM INDEX v_idx
    ON ks.v_table (row_v)
    USING 'StorageAttachedIndex';


/////////////////////////////
// SELECT WORKS (empty table)
/////////////////////////////

SELECT * FROM ks.v_table
    WHERE metadata['map_k'] = 'map_v'
        AND pk = 0
    ORDER BY row_v ANN OF [0.1, 0.2]
    LIMIT 4;


//////////////
// ADD ONE ROW
//////////////

INSERT INTO ks.v_table (pk, metadata, row_v)
VALUES
    (0, {'map_k': 'map_v'}, [0.11, 0.19]);



/////////////////////////////////////////////
// SELECT WORKS (table has queried partition)
/////////////////////////////////////////////

SELECT * FROM ks.v_table
    WHERE metadata['map_k'] = 'map_v'
        AND pk = 0
    ORDER BY row_v ANN OF [0.1, 0.2]
    LIMIT 4;


//////////////////////////////////
// ADD ONE ROW (another partition)
//////////////////////////////////

INSERT INTO ks.v_table (pk, metadata, row_v)
VALUES
    (10, {'map_k': 'map_v'}, [0.11, 0.19]);


/////////////////////////////////////////////////
// SELECT BREAKS (table gained another partition)
/////////////////////////////////////////////////

SELECT * FROM ks.v_table
    WHERE metadata['map_k'] = 'map_v'
        AND pk = 0
    ORDER BY row_v ANN OF [0.1, 0.2]
    LIMIT 4; 

The error has this appearance in CQL Console:

ReadFailure: Error from server: code=1300 [Replica(s) failed to execute read] message="Operation failed - received 0 responses and 1 failures: UNKNOWN from /172.17.0.2:7000" info={'consistency': 'ONE', 'required_responses': 1, 'received_responses': 0, 'failures': 1, 'error_code_map': {'172.17.0.2': '0x0000'}} 

And the Cassandra logs have this to say:

java.lang.NullPointerException: Cannot invoke "org.apache.cassandra.index.sai.iterators.KeyRangeIterator.skipTo(org.apache.cassandra.index.sai.utils.PrimaryKey)" because "this.nextIterator" is null 

 

Issue Links

• links to: GitHub Pull Request #3130