[
https://issues.apache.org/jira/browse/CASSANDRA-14100?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
ZhaoYang updated CASSANDRA-14100:
---------------------------------
Description:
In CASSANDRA-10657, value skipping is re-enabled to avoid reading unselected
columns from disk to reduce heap pressure.
In 3.11, ColumnFilter.builder#addAll() will initialize {{queriedBuilder}} as
{{empty}} even if the parameter is empty list.
In trunk, there is some changes on ColumnFilter.builder#addAll() that
ColumnFilter-queried will no longer be {{empty}} when no regular column is
selected, it's now {{null}} instead.
There is subtle difference between {{empty}} and {{null}} in ColumnFilter.
So {{ColumnFilter.fetchedColumnIsQueried}} will return true and
{{SerializationHeader.canSkipValue()}} will return {{false}} in trunk.
Probably we need to refactor ColumnFilter...
{code:title=reproduce}
// passed on 3.11, failed on trunk
createTable("CREATE TABLE %s (k1 int, v1 int, v2 int, PRIMARY KEY (k1))");
execute("INSERT INTO %s(k1,v1,v2) VALUES(1,1,1) USING TIMESTAMP 5");
flush();
Prepared prepared = QueryProcessor.prepareInternal("SELECT k1 FROM " +
keyspace() + "." + currentTable() + " WHERE k1=1");
CQLStatement cqlStatement = prepared.statement;
SelectStatement selectStatement = (SelectStatement) cqlStatement;
ColumnFilter columnFilter = selectStatement.queriedColumns();
// v1/v2 are fetched but not queried, so no need to read them from disk in
Cell.Serializer#deserialize() by checking
// SerializationHeader.canSkipValue()
assertFalse(columnFilter.fetchedColumnIsQueried(ColumnMetadata.regularColumn(keyspace(),
currentTable(), "v1", Int32Type.instance)));
assertFalse(columnFilter.fetchedColumnIsQueried(ColumnMetadata.regularColumn(keyspace(),
currentTable(), "v2", Int32Type.instance)));
{code}
was:
In CASSANDRA-10657, value skipping is re-enabled to avoid reading unselected
columns from disk to reduce heap pressure.
In 3.11, ColumnFilter.builder#addAll() will initialize {{queriedBuilder}} as
{{empty}} even if the parameter is empty list.
In trunk, CASSANDRA-7396 made some changes on ColumnFilter.builder#addAll()
that ColumnFilter-queried will no longer be {{empty}} when no regular column is
selected, it's now {{null}} instead.
There is subtle difference between {{empty}} and {{null}} in ColumnFilter.
So {{ColumnFilter.fetchedColumnIsQueried}} will return true and
{{SerializationHeader.canSkipValue()}} will return {{false}} in trunk.
Probably we need to refactor ColumnFilter...
{code:title=reproduce}
// passed on 3.11, failed on trunk
createTable("CREATE TABLE %s (k1 int, v1 int, v2 int, PRIMARY KEY (k1))");
execute("INSERT INTO %s(k1,v1,v2) VALUES(1,1,1) USING TIMESTAMP 5");
flush();
Prepared prepared = QueryProcessor.prepareInternal("SELECT k1 FROM " +
keyspace() + "." + currentTable() + " WHERE k1=1");
CQLStatement cqlStatement = prepared.statement;
SelectStatement selectStatement = (SelectStatement) cqlStatement;
ColumnFilter columnFilter = selectStatement.queriedColumns();
// v1/v2 are fetched but not queried, so no need to read them from disk in
Cell.Serializer#deserialize() by checking
// SerializationHeader.canSkipValue()
assertFalse(columnFilter.fetchedColumnIsQueried(ColumnMetadata.regularColumn(keyspace(),
currentTable(), "v1", Int32Type.instance)));
assertFalse(columnFilter.fetchedColumnIsQueried(ColumnMetadata.regularColumn(keyspace(),
currentTable(), "v2", Int32Type.instance)));
{code}
> proper value skipping when only pk columns are selected
> -------------------------------------------------------
>
> Key: CASSANDRA-14100
> URL: https://issues.apache.org/jira/browse/CASSANDRA-14100
> Project: Cassandra
> Issue Type: Bug
> Components: Local Write-Read Paths
> Reporter: ZhaoYang
> Priority: Minor
> Fix For: 4.x
>
>
> In CASSANDRA-10657, value skipping is re-enabled to avoid reading unselected
> columns from disk to reduce heap pressure.
> In 3.11, ColumnFilter.builder#addAll() will initialize {{queriedBuilder}} as
> {{empty}} even if the parameter is empty list.
> In trunk, there is some changes on ColumnFilter.builder#addAll() that
> ColumnFilter-queried will no longer be {{empty}} when no regular column is
> selected, it's now {{null}} instead.
> There is subtle difference between {{empty}} and {{null}} in ColumnFilter.
> So {{ColumnFilter.fetchedColumnIsQueried}} will return true and
> {{SerializationHeader.canSkipValue()}} will return {{false}} in trunk.
> Probably we need to refactor ColumnFilter...
> {code:title=reproduce}
> // passed on 3.11, failed on trunk
> createTable("CREATE TABLE %s (k1 int, v1 int, v2 int, PRIMARY KEY (k1))");
> execute("INSERT INTO %s(k1,v1,v2) VALUES(1,1,1) USING TIMESTAMP 5");
> flush();
> Prepared prepared = QueryProcessor.prepareInternal("SELECT k1 FROM " +
> keyspace() + "." + currentTable() + " WHERE k1=1");
> CQLStatement cqlStatement = prepared.statement;
> SelectStatement selectStatement = (SelectStatement) cqlStatement;
> ColumnFilter columnFilter = selectStatement.queriedColumns();
> // v1/v2 are fetched but not queried, so no need to read them from disk in
> Cell.Serializer#deserialize() by checking
> // SerializationHeader.canSkipValue()
> assertFalse(columnFilter.fetchedColumnIsQueried(ColumnMetadata.regularColumn(keyspace(),
> currentTable(), "v1", Int32Type.instance)));
> assertFalse(columnFilter.fetchedColumnIsQueried(ColumnMetadata.regularColumn(keyspace(),
> currentTable(), "v2", Int32Type.instance)));
> {code}
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
