aether/CLAUDE.md

Aether

Distributed actor system with event sourcing for Go, powered by NATS.

Organization Context

This repo is part of Flowmade. See:

• Organization manifesto - who we are, what we believe
• Repository map - how this fits in the bigger picture
• Vision - what this specific product does

Setup

git clone git@git.flowmade.one:flowmade-one/aether.git
cd aether
go mod download

Requires NATS server for integration tests:

# Install NATS
brew install nats-server

# Run with JetStream enabled
nats-server -js

Project Structure

aether/
├── event.go           # Event, ActorSnapshot, EventStore interface
├── eventbus.go        # EventBus, EventBroadcaster interface
├── nats_eventbus.go   # NATSEventBus - cross-node event broadcasting
├── store/
│   ├── memory.go      # InMemoryEventStore (testing)
│   └── jetstream.go   # JetStreamEventStore (production)
├── cluster/
│   ├── manager.go     # ClusterManager
│   ├── discovery.go   # NodeDiscovery
│   ├── hashring.go    # ConsistentHashRing
│   ├── shard.go       # ShardManager
│   ├── leader.go      # LeaderElection
│   └── types.go       # Cluster types
└── model/
    └── model.go       # EventStorming model types

Development

make build    # Build the library
make test     # Run tests
make lint     # Run linters

Architecture

Event Sourcing

Events are the source of truth. State is derived by replaying events.

// Create an event
event := &aether.Event{
    ID:        uuid.New().String(),
    EventType: "OrderPlaced",
    ActorID:   "order-123",
    Version:   1,
    Data:      map[string]interface{}{"total": 100.00},
    Timestamp: time.Now(),
}

// Persist to event store
store.SaveEvent(event)

// Replay events to rebuild state
events, _ := store.GetEvents("order-123", 0)

Event Versioning

Events for each actor must have monotonically increasing versions. This ensures event stream integrity and enables optimistic concurrency control.

Version Semantics

• Each actor has an independent version sequence
• Version must be strictly greater than the current latest version
• For new actors (no events), the first event must have version > 0
• Non-consecutive versions are allowed (gaps are permitted)

Optimistic Concurrency Pattern

// 1. Get current version
currentVersion, _ := store.GetLatestVersion("order-123")

// 2. Create event with next version
event := &aether.Event{
    ID:        uuid.New().String(),
    EventType: "OrderUpdated",
    ActorID:   "order-123",
    Version:   currentVersion + 1,
    Data:      map[string]interface{}{"status": "shipped"},
    Timestamp: time.Now(),
}

// 3. Attempt to save
err := store.SaveEvent(event)
if errors.Is(err, aether.ErrVersionConflict) {
    // Another writer won - reload and retry if appropriate
    var versionErr *aether.VersionConflictError
    errors.As(err, &versionErr)
    log.Printf("Conflict: actor %s has version %d, attempted %d",
        versionErr.ActorID, versionErr.CurrentVersion, versionErr.AttemptedVersion)
}

Error Types

• ErrVersionConflict - Sentinel error for version conflicts (use with errors.Is)
• VersionConflictError - Detailed error with ActorID, CurrentVersion, and AttemptedVersion

Version Cache Invalidation

The JetStreamEventStore maintains an in-memory cache of actor versions to optimize repeated version checks during optimistic concurrency control. The cache is designed to handle multi-store scenarios where external processes may write to the same JetStream stream:

• Cache hits: Cached version is returned immediately for performance
• Cache misses: If no cached version exists, JetStream is queried and cached
• External writes: If GetLatestVersion detects a version newer than cached, the cache is invalidated and fresh data is fetched next time

This strategy ensures data consistency even in scenarios with external writers while maintaining excellent performance for the single-writer case (where only Aether owns the stream).

Implementation detail: The cache is invalidated by deleting the entry, forcing a fresh fetch from JetStream on the next version check for that actor. This is safe because:

1. SaveEvent uses getLatestVersionLocked() which checks JetStream on cache miss
2. GetLatestVersion always fetches fresh data and detects stale cache entries
3. Subsequent checks will fetch from JetStream until the cache is repopulated

Namespace Isolation

Namespaces provide logical boundaries for events and subscriptions.

Event Bus Namespaces

The event bus supports namespace-scoped pub/sub:

// Subscribe to events in a namespace
ch := eventBus.Subscribe("tenant-abc")

// Events are isolated per namespace
eventBus.Publish("tenant-abc", event)  // Only tenant-abc subscribers see this

Namespace-Scoped Event Stores

JetStreamEventStore supports optional namespace prefixes for complete storage isolation:

// Create a namespaced event store (convenience function)
store, err := store.NewJetStreamEventStoreWithNamespace(natsConn, "events", "tenant-abc")

// Or configure via JetStreamConfig
config := store.JetStreamConfig{
    Namespace:       "tenant-abc",
    StreamRetention: 30 * 24 * time.Hour,
    ReplicaCount:    3,
}
store, err := store.NewJetStreamEventStoreWithConfig(natsConn, "events", config)

// The actual stream name becomes "tenant-abc_events"
// Events from one namespace cannot be read from another namespace's store

Namespace isolation at the storage level ensures:

• Complete data isolation: Events stored with one namespace prefix are invisible to stores with different namespaces
• Backward compatibility: Empty namespace (default) works exactly as before
• Safe characters: Namespace names are sanitized (spaces, dots, wildcards become underscores)

Use namespace-scoped stores when you need strong isolation guarantees at the persistence layer, such as:

• Multi-tenant deployments where tenant data must be completely separated
• Logical boundaries between different domains or bounded contexts
• Test isolation in integration tests

Clustering

Aether handles node discovery, leader election, and shard distribution:

// Create cluster manager
manager := cluster.NewClusterManager(natsConn, nodeID)

// Join cluster
manager.Start()

// Leader election happens automatically
if manager.IsLeader() {
    // Coordinate shard assignments
}

Key Patterns

• Events are immutable - Never modify, only append
• Versions are monotonic - Each event must have version > previous for same actor
• Snapshots for performance - Periodically snapshot state to avoid full replay
• Namespaces for isolation - Not multi-tenancy, just logical boundaries
• NATS for everything - Events, pub/sub, clustering all use NATS