[
https://issues.apache.org/jira/browse/CASSANDRA-14638?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16582078#comment-16582078
]
Sylvain Lebresne commented on CASSANDRA-14638:
----------------------------------------------
I agree both that this is unfortunate, and that I don't see any reasonable
backward compatible option. And I'm also in favor or restoring the pre-3.0
order, partly because this is indeed the order we've used for the longest, but
mainly because it doesn't make sense to have inconsistent ordering for static
and normal columns.
I would also argue that relying that strongly on the order of columns of a
{{SELECT *}} is a bad idea in the first place, and to the best of my knowledge,
no official documentation has ever pretended that the order was guaranteed
(don't get me wrong, this is a weak argument in that our documentation are
historically bad and if we were to guarantee only what is documented, we
wouldn't be guaranteeing much; still, it would clearly have been worth if we
had ever documented the order as guaranteed), so hopefully this won't break too
many users. In fact, on top of the big fat NEWS entry already mentioned, I'd be
in favor of updating the documentation to explicitly mention that the order of
columns is not specified and not guaranteed to be stable, and should thus not
be relied on.
> 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
> ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
>
> Key: CASSANDRA-14638
> URL: https://issues.apache.org/jira/browse/CASSANDRA-14638
> Project: Cassandra
> Issue Type: Bug
> Environment: Single C* node ccm cluster upgraded from C* 2.1.20 to
> 3.0.17
> Reporter: Andy Tolbert
> Assignee: Aleksey Yeschenko
> Priority: Major
> Fix For: 3.0.x, 3.11.x, 4.0.x
>
>
> 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:
>
> {noformat}
> 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');{noformat}
>
> When querying with 2.1 you'll observe the following order via cqlsh:
> {noformat}
> p0 | p1 | m | u | t
> ----+----+--------------+---+---
> p0 | p1 | {'m0': 'm1'} | u | t{noformat}
>
> With 3.0, observe that u and m are transposed:
>
> {noformat}
> p0 | p1 | u | m | t
> ----+----+---+--------------+---
> p0 | p1 | u | {'m0': 'm1'} | t{noformat}
>
>
> {code:java}
> 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();
> }
> }
> }
> {code}
> 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:
>
> {noformat}
> 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{noformat}
>
> With 3.x:
>
> {noformat}
> 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{noformat}
>
>
>
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org
For additional commands, e-mail: commits-h...@cassandra.apache.org
