[
https://issues.apache.org/jira/browse/CASSANDRA-6426?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Aleksey Yeschenko resolved CASSANDRA-6426.
------------------------------------------
Resolution: Not A Problem
In a batch, your statements with the same partition key aren't executed in any
order, they are executed together, 'merged' with one another, and the merge
follows the standard Cassandra resolution rules. And since you haven't
specified explicit timestamps for your UPDATE and INSERT, the tombstone ('null'
value) will always override any other column, no matter what order you add them
in to the batch.
You should either specify the explicit timestamps (and make sure they increase
for every statement) or not update AND delete any columns using the same
timestamp in a single batch (DELETE or UPDATE SET NULL will always win given a
tie in the timestamp).
> Lost update in Prepared statement batch when a column is inserted with null
> value in the same batch just before
> ---------------------------------------------------------------------------------------------------------------
>
> Key: CASSANDRA-6426
> URL: https://issues.apache.org/jira/browse/CASSANDRA-6426
> Project: Cassandra
> Issue Type: Bug
> Components: Core
> Environment: Cassandra Server 2.0.3
> Java Driver Core 2.0.0-rc1
> Reporter: DOAN DuyHai
>
> {panel:title=Test environment}
> Cassandra Server *2.0.3*
> Java Driver Core *2.0.0-rc1*
> {panel}
> While implementing batched prepared statements for Achilles, I ran into a
> very annoying bug.
> {code:title=FailingUnitTest.java|borderStyle=solid}
> //Given
> Long id = RandomUtils.nextLong();
> session.execute("CREATE TABLE test(id bigint, name text,label text, PRIMARY
> KEY(id))");
> PreparedStatement insertPS = session.prepare("INSERT INTO test(id,name,label)
> VALUES (?,?,?)");
> PreparedStatement updatePS = session.prepare("UPDATE test SET label=? WHERE
> id=?");
> // Notice the "label" column is inserted first with null
> BoundStatement insertBS = insertPS.bind(id, "myName", null);
> // Then "label" is updated to "myLabel"
> BoundStatement updateBS = updatePS.bind("myLabel", id);
> //When
> BatchStatement batch = new BatchStatement();
> batch.add(insertBS);
> batch.add(updateBS);
> session.execute(batch);
> //Then
> Statement statement = new SimpleStatement("SELECT * from test where id=" +
> id);
> Row row = session.execute(statement).one();
> // Assertion FAIL, "label" is NULL
> assertThat(row.getString("label")).isEqualTo("myLabel");
> {code}
> The above code always fails.
> Even if I switch the order of the bound statements, e.g. *updateBS* added to
> the batch before *insertBS*, the test still fails.
> {code:title=WorkingUnitTest.java|borderStyle=solid}
> //Given
> Long id = RandomUtils.nextLong();
> session.execute("CREATE TABLE test(id bigint, name text,label text, PRIMARY
> KEY(id))");
> PreparedStatement insertPS = session.prepare("INSERT INTO test(id,name)
> VALUES (?,?)");
> PreparedStatement updatePS = session.prepare("UPDATE test SET label=? WHERE
> id=?");
> // Notice the "label" column is removed from the insert statement
> BoundStatement insertBS = insertPS.bind(id, "myName");
> BoundStatement updateBS = updatePS.bind("myLabel", id);
> //When
> BatchStatement batch = new BatchStatement();
> batch.add(updateBS);
> batch.add(insertBS);
> session.execute(batch);
> //Then
> Statement statement = new SimpleStatement("SELECT * from test where id=" +
> id);
> Row row = session.execute(statement).one();
> // Assertion SUCCEEDS here
> assertThat(row.getString("label")).isEqualTo("myLabel");
> {code}
> Only removing column "label" from the first insert bound statement can make
> the test succeed. This is pretty annoying.
> The rationale for inserting null values is that in Achilles I prepare
> generic insert statements for all entities and setting NULL when the field is
> not filled at runtime.
> Currently, there is no guarantee for batch of prepared statement to be
> executed in the order they are declared.
--
This message was sent by Atlassian JIRA
(v6.1#6144)
