Github user aweisberg commented on a diff in the pull request:
https://github.com/apache/cassandra/pull/224#discussion_r187429139
--- Diff:
src/java/org/apache/cassandra/cql3/statements/AlterKeyspaceStatement.java ---
@@ -96,7 +98,35 @@ private void warnIfIncreasingRF(KeyspaceMetadata ksm,
KeyspaceParams params)
StorageService.instance.getTokenMetadata(),
DatabaseDescriptor.getEndpointSnitch(),
params.replication.options);
- if (newStrategy.getReplicationFactor() >
oldStrategy.getReplicationFactor())
+
+ validateTransientReplication(oldStrategy, newStrategy);
+ warnIfIncreasingRF(oldStrategy, newStrategy);
+ }
+
+ private void validateTransientReplication(AbstractReplicationStrategy
oldStrategy, AbstractReplicationStrategy newStrategy)
+ {
+ if (oldStrategy.getReplicationFactor().trans == 0 &&
newStrategy.getReplicationFactor().trans > 0)
+ {
+ Keyspace ks = Keyspace.open(keyspace());
+ for (ColumnFamilyStore cfs: ks.getColumnFamilyStores())
+ {
+ if (cfs.viewManager.hasViews())
+ {
+ throw new ConfigurationException("Cannot use transient
replication on keyspaces using materialized views");
+ }
+
+ if (cfs.indexManager.hasIndexes())
+ {
+ throw new ConfigurationException("Cannot use transient
replication on keyspaces using secondary indexes");
+ }
+ }
+
+ }
+ }
--- End diff --
I think we can allow people to increase # of transient replicas (they make
it real with nodetool cleanup or repair), but how does decreasing number of
transient replicas work safely? I think we should disallow it until we can
articulate how and why it is safe. There has to be a temporary pending state
where a node is transitioning from transient to full where it receives writes
but not reads because it can't correctly service reads as a full replcia.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
