Skip to content

ReplicaNodeIdentity.md

Latest commit

 

History

History
191 lines (155 loc) · 9.08 KB

ReplicaNodeIdentity.md

File metadata and controls

191 lines (155 loc) · 9.08 KB
Raw

Replica-consistent NodeIds

OPC UA Part 4, 6.6.2.2 requires a RedundantServerSet to expose the same application NodeIds on every replica. Matching a namespace URI after client-side translation is not sufficient: the namespace index, identifier type and identifier value must agree. Built-in server diagnostics and configuration remain local.

Opc.Ua.Redundancy.Server supplies ReplicaNodeIdFactory and UseReplicaNodeIdentity for both active/passive and active/active address spaces. The module wraps the standard DefaultNodeIdFactory; it does not introduce a second hashing algorithm or a persistent logical-key-to-NodeId database.

Contents

Configure the whole replica set

Every replica uses the same replica-set ID, ordered namespace list and assignment mode. Its own ApplicationUri and election/gossip replica ID remain distinct.

serverBuilder
    .UseReplicaNodeIdentity(
        "production-line",
        ["urn:example:machine-model", "urn:example:machine-instances"],
        NodeIdAssignmentMode.Numeric,
        writerAssignedIds: true)
    .UseDistributedAddressSpace(options =>
    {
        options.UseLeaderElection = true;
        options.NodeId = localReplicaId;
    });

Supply the shared store and record protector using the normal HA composition. For active/active, leave writerAssignedIds false:

serverBuilder
    .UseReplicaNodeIdentity(
        "production-line",
        ["urn:example:machine-model", "urn:example:machine-instances"])
    .UseActiveActiveRedundancy(options =>
    {
        options.ReplicaId = replicaId;
        options.GossipPort = 4840;
        options.Tls = mutualTlsOptions;
        options.AddPeer(peerEndpoint);
    });

In a real active/active configuration, set GossipPort, Tls, and peer endpoints on ActiveActiveRedundancyOptions, as shown in the active/active example. Do not combine active/passive and active/active address-space modules on one server.

The configured slots are:

Index Meaning
0 Standard OPC UA namespace
1 This server's local ApplicationUri
2 onwards Shared namespace URIs, in the configured order

The layout is installed before node-manager constructors run. Existing entries are never renumbered; a conflicting slot fails startup. Declare shared model and instance namespaces before creating nodes, including namespaces needed by their types, references and NodeId-valued data. New shared namespaces require an explicit layout revision, not a different runtime append order.

Validation checks:

  • Namespace-bearing type definitions, data types, modelling rules, references, and role permissions.
  • NodeId, ExpandedNodeId, and QualifiedName values inside arrays, matrices, DataValues, and encodeable structures, including method arguments.

The validator visits structured values through their normal encoding contract, without reflection. Register a codec for an opaque ExtensionObject before using it as shared data; the validator cannot certify an unknown body as namespace-safe.

Before publishing a replacement manager, the runtime validates its ownership composition and hydrates retained identities into it. After publication, the module rebinds capture and hydration to the new manager instances. Active/passive reconfiguration retains the shared store and election. Active/active reconfiguration retains the gossip transport and CRDT map.

If preparation fails, the runtime restores bindings for the manager composition that the lifecycle retained. Replication therefore does not remain attached only to retired managers.

Both server startup modules reject missing identity configuration for replicated operation. Process-local active/passive experimentation and ordinary standalone servers retain their existing behavior.

For non-DI hosting, assign a ReplicaNodeIdFactory to StandardServer.NodeIdFactory before starting the server. Supply the writer election, shared state store and record protector to its constructor when using writer-assigned IDs. The server awaits the same early initialization, and the address-space startup task validates the same contract.

Creating, importing and hydrating nodes

Numeric, String, Guid and Opaque modes derive named identities through the standard canonical-path implementation. Numeric remains the default. A shared path must not depend on an unregistered namespace or the local ApplicationUri. Independently created entities need a stable application key represented by their path or an explicit NodeId. Repeated browse names are not a substitute for distinct stable keys: supply distinct explicit identifiers when a path is ambiguous.

Counter mode and unnamed-node fallback do not provide independent identity. With writerAssignedIds: true, only the elected writer may register a locally allocated counter ID in the shared graph. Standbys consume the writer's assigned IDs unchanged. Startup reserves live IDs and retained tombstones before writer allocation; hydration also reserves incoming IDs. There is no synchronous storage or network call in New.

Valid supplied IDs and hydrated child IDs are retained, not allocated again. The shared registration policy checks collisions before indexing, including collisions with explicit IDs. Hashed shared allocations keep collision detection enabled in Release as well as Debug. Rebased factory views and supported factory replacement preserve the policy; an unrelated custom factory cannot replace it. Custom application identity schemes must supply their stable IDs explicitly.

Transient event instances and their generated fields use the server-local namespace; they are not independently registered shared entities. Built-in diagnostics/configuration and namespace-zero infrastructure are excluded from address-space replication. WoT shared materialization retains the configured stable factory rather than forcing independent shared counters.

Stored contracts and peer admission

The versioned descriptor contains the replica-set ID, canonical-path version, assignment mode, writer/independent policy and ordered namespace layout.

Active/passive stores bind it to the protected strong key election/addressspace-identity/v1. Fluent registration contributes that key even with custom hybrid routing. A new authoritative store can initialize it atomically; competing initializers must accept the winning descriptor. Mismatched, corrupt or unbound legacy state fails without overwriting that state.

An empty CRDT scan is not proof of an empty store. For a verified new hybrid deployment, provision the contract on its fresh linearizable backend before attaching the also-new replicated payload backend:

var identity = new ReplicaNodeIdFactory(
    "production-line",
    ["urn:example:machine-model", "urn:example:machine-instances"]);
await identity.InitializeNewStoreAsync(newSharedRaftStore, recordProtector, ct);

Use the same writer policy in the provisioning factory and the hosted factory: if the host enables writer-assigned IDs, pass its election to the provisioning factory too. This administrative operation cannot inspect disconnected replicas and is not a legacy-state adoption API. Never use it to certify an existing unknown payload store. Export and validate an existing layout offline before rebuilding or migrating; there is no automatic live renumbering or store wipe.

Active/active frames carry the descriptor over the existing gossip transport and compare it before CRDT merge. An incompatible frame is rejected, does not advance the successful-apply counter, and invalidates local identity admission. The configured service-level wiring keeps that replica at maintenance level 0; correct the configuration and restart it. Configure authenticated gossip as usual: the descriptor is compatibility metadata, not a replacement for transport security. All opted-in ownership partitions use one address-space transport per replica, so multiple node managers do not each bind the same listening port.

Fixed identity does not make eventual values linearizable, infer missing-node deletion from incomplete scans, or relax the existing HA coordination rules.

Sample client proof

The RedundantServer sample reserves its shared namespace through the module and exposes factory-created HighAvailability/FactoryAssigned/Value and Target nodes. Target contains the exact NodeId of Value.

Run the RedundantClient sample with --identity, wait for its cached-ID message, then stop the serving replica. The client verifies a different server through the standard ServerArray, reads the saved IDs without rebrowsing or remapping, and creates a monitored item using the saved value ID on the surviving replica.