* send terminationMessage to singleton when leaving last, #21592
* When leaving last node, i.e. no newOldestOption, the manager was
just stopped. The change is to send the terminationMessage also
in this case and wait until the singleton actor is terminated
before stopping the manager.
* Also changed so that the singleton is stopped immediately when
cluster has been terminated when last node is leaving, i.e.
no newOldestOption. Previously it retried until maxTakeOverRetries
before stopping.
* More comprehensive test of this scenario in ClusterSingletonManagerLeaveSpec
* increase test timeout
* When the test fails the node is removed from the membership
twice, which triggers two OldestChanged cycles, but in
the 2.4.9 change https://github.com/akka/akka/pull/21152/files#diff-f0ae95c926a050aecf45dba3e08d1c77L669
the singleton manager always goes to End (stop) when it has been Oldest
* This fix restores the previous behavior for this scenario
* track nodes by UniqueAddress in Cluster Singleton, #20942
* reply with HandOverDone from new incarnation, #20942
* confirm as terminated immediately when new incarnation joins, #20942 instead of waiting for failure detector to mark it as unreachable this will speed-up removal when restarting cluster node with same hostname:port
* ignore PubSub Status message from unknown node, #20846
Reproducer:
1. old cluster of node1, node2 and node3
2. shutdown node3 and start it again with same host:port, let it
join itself and not the old cluster
3. node1 and node2 will continue to gossip to the node3 address and
Status message is accepted and replied to (Delta is ignored from
unknown node)
Solution:
* ignore status message from unknown node
* also added a reply flag in the Status message to break the
back-and-forth replies in case the deltas are not accepted,
this is not needed for fixing this bug, but it adds an extra
level of safety
This ensures that gracefully leaving nodes (which would terminate after
their own removed event) are already unregistered on all other pubsub
nodes, before termination.
* Provide shorter aliases for the ActorRefProviders #20649
* Use the new actorefprovider aliases throughout code and docs
* Cleaner alias replacement logic
Allows the cluster client and its receptionist to be observable in terms of contact points becoming available and client heartbeats. Furthermore a query API for requesting the current state has been provided.
GZIPInputStream uses Inflater internally (so also native zlib). Inflater frees up memory only on explicit call to end() or during finalization (finalize() contains only call to end()), so GZIPInputStream should always be explicitly closed.
As native libraries are used a non-scalaish try-finally is used to avoid off-heap memory leak for GZIPInputStream and GZIPOutputStream in case of exceptions.
* Added reusable PerGroupingBuffer trait for pubsub implementation
* Moved mkKey methods to the Internal object
* Introduced passivate-like protocol between DistributedPubSubMediator/Topic and Topic/Group actors, contained in ChildActorTerminationProtocol messages.
* the reported issue is fixed by the immediate leaderActions
(moving to Up) when joining the first node to itself
* the other changes are precautions just in case
* In 2.4 we derive the number of hand-over/take-over retries from
the removal margin, but we decided to set that to 0 by default, since
it is intended for network partition scenarios. maxTakeOverRetries
became 1. So there must be also be a min number of retries property.
* The test failed for the leaving scenario because the singleton
instance was stopped hard without sending the terminationMessage when
the maxTakeOverRetries was exceeded.
For manual downing it is not needed. For auto-down it doesn't add any extra safety, since that
is not handling network partitions anyway.
The setting is still useful if you implement downing strategies that handle network partitions,
e.g. by keeping the larger side of the partition and shutting down the smaller side.
- created new subproject akka-protobuf (and added COPYING and LICENSE)
- renamed com.google.protobuf -> akka.protobuf everywhere
- also added such renaming step to the results of protoc compilation in
project/Protobuf.scala
- had to include transcriptions of Netty’s ProtobufEncoder/Decoder to
make multi-node-testkit compile again
