Github user belliottsmith commented on a diff in the pull request:
https://github.com/apache/cassandra/pull/270#discussion_r218488201
--- Diff: src/java/org/apache/cassandra/locator/ReplicaLayout.java ---
@@ -275,7 +275,24 @@ public ForTokenWrite(EndpointsForToken natural,
EndpointsForToken pending, Endpo
*/
private static <E extends Endpoints<E>> E
resolveWriteConflictsInNatural(E natural, E pending)
{
- return natural.filter(r -> !r.isTransient() ||
!pending.contains(r.endpoint(), true));
+ ReplicaCollection.Mutable<E> resolved =
natural.newMutable(natural.size());
+ for (Replica replica : natural)
+ {
+ // always prefer the full natural replica, if there is a
conflict
+ if (!replica.isFull())
+ {
+ Replica conflict =
pending.byEndpoint().get(replica.endpoint());
+ if (conflict != null)
+ {
+ // If we have any pending transient->full movement, we
need to move the full replica to our 'natural' bucket
+ // to avoid corrupting our count
+ if (conflict.isFull()) resolved.add(new
Replica(replica.endpoint(), replica.range(), true));
--- End diff --
No, you're right, good catch. Even if they are a different range, either
should be acceptable for this query (by definition). I'll push an update.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
