Service Fabric Reliable Collections that report storage load metrics.

Metric Reliable Collections are wrappers around Reliable Collections that automatically report memory and disk usage load metrics for automated resource balancing of stateful services based on the amount of data they're storing. The load reported is exactly the amount of data stored in Reliable Collections in each replica of a service that uses Metric Reliable Collections. It does not report total memory and disk used by each replica. Both primary and active secondary replicas report load.

1. Simply replace your ReliableStateManager instance with a MetricReliableStateManager when you create your service class. As long as your service works only with the Reliable Collection interfaces and not the concrete classes, no other changes are required to your service code. You must provide a custom serializer. A JSON serializer is included to get started - it's not the fastest or the most efficient, but it should work with just about any data type.

   internal static class Program
   {
       private static void Main()
       {
           ServiceRuntime.RegisterServiceAsync(
               "LoadGenServiceType",
               context => new LoadGenService(
                   context,
                   new MetricReliableStateManager(
                       context,
                       new JsonReliableStateSerializerResolver(),
                       new MetricConfiguration("MemoryKB", DataSizeUnits.Kilobytes, "DiskKB", DataSizeUnits.Kilobytes, TimeSpan.FromSeconds(10), TimeSpan.FromSeconds(5)))))
               .GetAwaiter()
               .GetResult();
               
           // Prevents this host process from terminating so services keep running.
           Thread.Sleep(Timeout.Infinite);
       }
   }

2. Set up default metric values for your service instances that match the names provided in MetricConfiguration.

Default Services in Visual Studio:

   <DefaultServices>
     <Service Name="LoadGenService">
        <StatefulService ServiceTypeName="LoadGenServiceType" TargetReplicaSetSize="3" MinReplicaSetSize="2">
           <UniformInt64Partition PartitionCount="10" LowKey="0" HighKey="10" />
           <LoadMetrics>
              <LoadMetric Name="MemoryKB" />
              <LoadMetric Name="DiskKB" />
           </LoadMetrics>
        </StatefulService>
     </Service>
  </DefaultServices>

PowerShell:

New-ServiceFabricService -ApplicationName "fabric:/LoadMetrics" -ServiceName "fabric:/LoadMetrics/LoadGenService" -ServiceTypeName "LoadGenServiceType" -Stateful -PartitionSchemeUniformInt64 -PartitionCount 10 -LowKey 0 -HighKey 10 -HasPersistedState -TargetReplicaSetSize 3 -MinReplicaSetSize 3 -Metric @("MemoryKB,High,0,1”, "DiskKB,High,0,1”)

3. Set up Node capacities for memory and disk with the same names provided in MetricConfiguration and in the service instances;

Cluster Manifest:

   <NodeTypes>
       <NodeType Name="NodeType0">
           <Endpoints>
               <ClientConnectionEndpoint Port="19000" />
               <LeaseDriverEndpoint Port="19001" />
               <ClusterConnectionEndpoint Port="19002" />
               <HttpGatewayEndpoint Port="19080" Protocol="http" />
               <ServiceConnectionEndpoint Port="19006" />
               <HttpApplicationGatewayEndpoint Port="19081" Protocol="http" />
               <ApplicationEndpoints StartPort="30001" EndPort="31000" />
           </Endpoints>
   		<Capacities>
               <Capacity Name="MemoryKB" Value="1048576"/>
               <Capacity Name="DiskKB" Value="104857600"/>
           </Capacities>
       </NodeType>
      ...
   </NodeTypes>

Services that use Reliable Collections program against the Reliable Collection interfaces. Metric Reliable Collections are implementations of those interfaces. For services that are written against the interfaces, no change to service code should be required to use Metric Reliable Collections.

At the heart is MetricReliableStateManager, an implementation of IReliableStateManager. It manages MetricReliableCollections, implementations of IReliableCollections. MetricReliableStateManager is a wrapper around the built-in ReliableStateManager that creates Metric wrappers around Reliable Collections. The Reliable Collections that are created store byte arrays, and the Metric wrappers use your serializer to convert your strongly-typed objects to byte arrays. The memory and disk usage that gets reported is the size of the byte arrays, which is directly affected by your serialization. This allows the Metric classes to quickly and accurately compute usage by totalling up the number of bytes stored in the collections without walking object graphs to compute object size. This also allows MetricReliableStateManager to define a flexible custom serialization interface.

Still a work-in-progress but ready for a test run.

• Background reading on cluster capacity: https://azure.microsoft.com/en-us/documentation/articles/service-fabric-cluster-resource-manager-cluster-description/#cluster-capacity

• Background reading on load metrics: https://azure.microsoft.com/en-us/documentation/articles/service-fabric-cluster-resource-manager-metrics/