[CASSANDRA-14638] Column result order can change in 'SELECT *' results when upgrading from 2.1 to 3.0 causing response corruption for queries using prepared statements when static columns are used

• Type: Bug
• Status: Resolved
• Priority: Normal
• Resolution: Fixed
• Fix Version/s: 3.0.18, 3.11.4, 4.0-alpha1, 4.0
• Component/s: CQL/Interpreter

Environment

Single C* node ccm cluster upgraded from C* 2.1.20 to 3.0.17

• Bug Category: Correctness - Transient Incorrect Response
• Severity: Normal
• Complexity: Normal
• Discovered By: User Report
• Since Version: 3.0.0
• Assignee: Aleksey Yeschenko
• Reporter: Andy Tolbert

Authors

Aleksey Yeschenko

Reviewers

Benedict Elliott Smith

• Created: 10/Aug/18 17:13
• Updated: 15/May/20 08:02
• Resolved: 16/Aug/18 14:50

Description

When performing an upgrade from C* 2.1.20 to 3.0.17 I observed that the order of columns returned from a 'SELECT *' query changes, particularly when static columns are involved.

This may not seem like that much of a problem, however if using Prepared Statements, any clients that remain connected during the upgrade may encounter issues consuming results from these queries, as data is reordered and the client not aware of it.  The result definition is sent in the original prepared statement response, so if order changes the client has no way of knowing (until C* 4.0 via CASSANDRA-10786) without re-preparing, which is non-trivial as most client drivers cache prepared statements.

This could lead to reading the wrong values for columns, which could result in some kind of deserialization exception or if the data types of the switched columns are compatible, the wrong values.  This happens even if the client attempts to retrieve a column value by name (i.e. row.getInt("colx")).

Unfortunately I don't think there is an easy fix for this.  If the order was changed back to the previous format, you risk issues for users upgrading from older 3.0 version.  I think it would be nice to add a note in the NEWS file in the 3.0 upgrade section that describes this issue, and how to work around it (specify all column names of interest explicitly in query).

Example schema and code to reproduce:

 

create keyspace ks with replication = {'class': 'SimpleStrategy', 'replication_factor': 1};

create table ks.tbl (p0 text,
  p1 text,
  m map<text, text> static,
  t text,
  u text static,
  primary key (p0, p1)
);

insert into ks.tbl (p0, p1, m, t, u) values ('p0', 'p1', { 'm0' : 'm1' }, 't', 'u');

 

When querying with 2.1 you'll observe the following order via cqlsh:

 p0 | p1 | m            | u | t
----+----+--------------+---+---
 p0 | p1 | {'m0': 'm1'} | u | t

 

With 3.0, observe that u and m are transposed:

 

 p0 | p1 | u | m            | t
----+----+---+--------------+---
 p0 | p1 | u | {'m0': 'm1'} | t

 

 

import com.datastax.driver.core.BoundStatement;
import com.datastax.driver.core.Cluster;
import com.datastax.driver.core.ColumnDefinitions;
import com.datastax.driver.core.PreparedStatement;
import com.datastax.driver.core.ResultSet;
import com.datastax.driver.core.Row;
import com.datastax.driver.core.Session;
import com.google.common.util.concurrent.Uninterruptibles;
import java.util.concurrent.TimeUnit;

public class LiveUpgradeTest {

  public static void main(String args[]) {
    Cluster cluster = Cluster.builder().addContactPoints("127.0.0.1").build();
    try {
      Session session = cluster.connect();
      PreparedStatement p = session.prepare("SELECT * from ks.tbl");

      BoundStatement bs = p.bind();

      // continually query every 30 seconds
      while (true) {
        try {
          ResultSet r = session.execute(bs);
          Row row = r.one();
          int i = 0;
          // iterate over the result metadata in order printing the
          // index, name, type, and length of the first row of data.
          for (ColumnDefinitions.Definition d : r.getColumnDefinitions()) {
            System.out.println(
                i++
                    + ": "
                    + d.getName()
                    + " -> "
                    + d.getType()
                    + " -> val = "
                    + row.getBytesUnsafe(d.getName()).array().length);
          }
        } catch (Throwable t) {
          t.printStackTrace();
        } finally {
          Uninterruptibles.sleepUninterruptibly(30, TimeUnit.SECONDS);
        }
      }
    } finally {
      cluster.close();
    }
  }
}

To reproduce, set up a cluster, the schema, and run this script.  Then upgrade the cluster to 3.0.17 (with ccm, ccm stop; ccm node1 setdir -v 3.0.17; ccm start works) and observe after the client is able to reconnect that the results are in a different order.  i.e.:

 

With 2.x:

 

0: p0 -> varchar -> val = 2
1: p1 -> varchar -> val = 2
2: m -> map<varchar, varchar> -> val = 16
3: u -> varchar -> val = 1
4: t -> varchar -> val = 1

 

With 3.x:

 

0: p0 -> varchar -> val = 2
1: p1 -> varchar -> val = 2
2: m -> map<varchar, varchar> -> val = 1
3: u -> varchar -> val = 16 (<-- the data for 'm' is now at index 3)
4: t -> varchar -> val = 1