Akka.NET v1.5.72 backports a non-blocking Cluster extension startup, test-infrastructure hardening, a Cluster.Sharding hand-over fix, and a batch of de-flaked specs from dev.

Starting with this release, Akka.Hosting ships from this repository. The Akka.Hosting, Akka.Remote.Hosting, Akka.Cluster.Hosting, Akka.Persistence.Hosting, Akka.Hosting.TestKit, and Akka.Hosting.TestKit.Xunit2 packages are now built, tested, and published alongside every Akka.NET release at the same version number.

Akka.Hosting

• Akka.Hosting moves into the akka.net repository - The six Hosting packages now live under src/contrib/hosting and share the Akka.NET version and release cadence. No package IDs, namespaces, or public APIs changed; bump Akka.Hosting.* to the same version as Akka.*. Hosting-only hotfix versions (such as 1.5.60.1) no longer exist: a Hosting fix ships in the next Akka.NET patch release. The documentation moved to getakka.net/articles/hosting. The akkadotnet/Akka.Hosting repository will be archived and keeps the full commit history and release notes for versions up to 1.5.71; the import was taken from commit cefe3c4c.
• Microsoft.Extensions minimum version moves to 10.0 - The Akka, Akka.DependencyInjection and all Akka.Hosting.* packages now require Microsoft.Extensions.* 10.0.0 or later (previously 6.0 for the core packages and 9.0 for Akka.Hosting). The 10.0 packages still ship netstandard2.0 assets, so applications on .NET Framework 4.8, .NET 6 and .NET 8 keep working; NuGet will lift the transitive Microsoft.Extensions.* references in those applications to 10.0.x. Akka.Hosting's OpenTelemetry dependency moves to 1.15.3 or later, which clears GHSA-g94r-2vxg-569j.

Akka.TestKit / Akka.TestKit.Xunit

• TestKit.Xunit: implement the async dispose chain - Akka.TestKit.Xunit.TestKit now implements InitializeAsync/DisposeAsync directly instead of relying on a derived class's own no-op overrides. Breaking change: a derived spec that already declares its own InitializeAsync/DisposeAsync must mark them override and call the base method, or the build fails with CS0114. Dispose(bool) no longer terminates the ActorSystem by itself; the public Dispose()/DisposeAsync() entry points do, after calling the base chain.
• Put the DotNetty batching override under akka.remote - The TestKit's write-batching override now lands under the key the transport actually reads. Every TestKit-hosted and multi-node test system runs with DotNetty write batching genuinely off for the first time.
• Add TestKitBase.ShutdownAsync - New public TestKitBase.ShutdownAsync overloads shut an ActorSystem down without blocking a thread pool worker.
• Eliminate the MultiNodeTestRunner conductor port race - The conductor node now binds port 0 and reports the real port over a stdout sentinel, instead of guessing a free port in advance, removing a race that could leave every other node waiting out a timeout against a conductor that never started. Adds the public ConductorBindException and ConductorPortSentinel types, and a 20-minute backstop that kills a node process that never exits on its own.

Akka.Streams

• Tcp Unbind completes immediately when idle - ServerBinding.Unbind() used to always wait out the subscription timeout even when no connection was pending. It now completes right away in that case, and still correctly waits when a connection was accepted but never subscribed to.

Akka.Cluster.Sharding

• Arm the remember-entities write timeout with updating-state-timeout - Remember-entities writes are now timed against updating-state-timeout (5 seconds by default), matching what reference.conf always documented, instead of waiting-for-state-timeout (2 seconds). Migration: if you raised waiting-for-state-timeout to give a slow remember-entities store more time, move that value to updating-state-timeout instead - it no longer affects the write path.
• A coordinator hand-over only stops the local ShardRegion when the node is leaving - Fixes #8583: a routine singleton hand-over during a rolling restart no longer stops a ShardRegion that isn't leaving the cluster, which used to leave that region permanently unusable for the life of the process.

Akka.DistributedData

• Fix Replicator.IsKnownNode to trust members first seen as Leaving/Exiting/Downed - A Replicator started after a member had already begun leaving no longer silently drops that member's writes and gossip.

Akka.Cluster

• Prevent removed cluster members from being resurrected by stale gossip - Gossip now carries UID-scoped removal tombstones and retains them for 24 hours by default (akka.cluster.prune-gossip-tombstones-after). The wire change is rolling-upgrade compatible: older nodes ignore the added protobuf field and upgraded nodes treat its absence as an empty tombstone set. Protection becomes fully effective after every node is upgraded; removals that occur while old nodes remain in the cluster are not guaranteed to retain their tombstones.
• Akka.Cluster: make Cluster extension startup non-blocking - The Cluster constructor no longer blocks on a 20 second ask, and akka.actor.creation-timeout is no longer consulted by Akka.Cluster at all (Akka.Cluster.Sharding still uses it to bound its own start-up asks). This removes the "Failed to startup Cluster" failure mode, which could shut down a healthy node whenever every thread the startup needed was parked - a small fork-join or other dedicated dispatcher pool, or a stock pool starved under load. Behavior note: Cluster.Get() now returns before core initialization completes, and the cluster's internal actor tree (/system/cluster/core, /system/cluster/core/daemon, /system/cluster/heartbeatReceiver) is created after Cluster.Get() returns rather than before it. All public cluster APIs are message-based, so typical code is unaffected; code that used Cluster.Get() as a "core is started" synchronization point, or that resolves one of those internal paths right after obtaining the extension, should await JoinAsync/RegisterOnMemberUp or a cluster event instead.
• Akka.Cluster: route user commands through /system/cluster for the extension's lifetime; LeaveAsync re-sends its Leave - All user-issued cluster commands (Subscribe, Join, Leave, Down, ...) now route through /system/cluster for the entire lifetime of the extension - two extra local mailbox hops on this low-traffic control plane - instead of switching to a direct core-daemon reference once startup resolves one, so back-to-back commands from one caller stay in order for good, not just during startup. A lost Leave no longer wedges every later LeaveAsync() for the life of the process. Behavior note: calling Leave()/LeaveAsync() again after the extension has already shut down now produces a dead-lettered ClusterUserAction.Leave - worth knowing about for dead-letter alerting or EventFilter.DeadLetter assertions.

Testing

• Bugfix5962Spec de-flaked for the non-blocking cluster startup - the spec's raw 1 second ResolveOne against /system/cluster/core/daemon/downingProvider raced the async cluster-extension init #8359 introduces above; replaced with a dilated retrying resolve, a dynamic port, and an async TaskCompletionSource member-up signal.
• De-flaked a large batch of Cluster, Cluster.Sharding, Cluster.Tools, DistributedData, Streams, and Remote specs, both single-process and multi-node, to remove timing-sensitive failures under load.

3 contributors since release 1.5.71

To see the full set of changes in Akka.NET v1.5.72, click here.