Distributed Orchestration

Current state

Cosmictron is currently a single-node runtime. All state, execution, and subscriptions live within one binary. This design provides:

• Deterministic event ordering without distributed consensus
• No network latency for in-process reducer calls
• Simple operational model — one process, one storage directory

For the majority of workloads (including large-scale AI agent deployments), a single well-provisioned Cosmictron instance handles thousands of concurrent sessions.

Cross-node orchestration (shipped)

While the storage engine is single-node, cross-node agent coordination is supported via:

Scoped isolation lifecycle

Each module runs in an isolated scope. Isolation boundaries can be managed programmatically:

// Pause a module scope (e.g., for maintenance)
IsolationScope::pause("module-name")?;

// Resume
IsolationScope::resume("module-name")?;

Rollout and restore

Controlled rollout of new module versions:

cosmictron-cli deploy my-agent.wasm --name my-agent --rollout 10%
# 10% of new sessions use the new version

cosmictron-cli deploy my-agent.wasm --name my-agent --rollout 100%
# Full rollout

cosmictron-cli rollback my-agent
# Restore previous version

External agent coordination

Cosmictron exposes a stable WebSocket + HTTP surface. Multiple Cosmictron instances can coordinate by reading from each other’s HTTP REST endpoints, using the event log as the coordination primitive.

Multi-region Coming Soon

Multi-region active-active deployment with automatic failover is planned. The design involves:

• Region-local primaries — writes committed locally first
• Async cross-region replication — event log replicated to other regions
• Conflict resolution — CRDT-based for agent state, strict ordering for financial events

There is no committed timeline for multi-region. It will not ship before a stable single-node product is proven in production.

Distributed Key Generation (DKG) Coming Soon

The current FROST threshold signing implementation requires key shares to be pre-distributed out-of-band. A built-in Distributed Key Generation (DKG) protocol that generates threshold keys without a trusted dealer is planned.

DKG is a prerequisite for fully trustless multi-party signing in multi-region deployments.

Caution

Do not plan production deployments on multi-region or DKG features. They are not shipped and do not have committed timelines.

Horizontal scaling today

Until multi-region ships, horizontal scaling options are:

1. Vertical scaling — Cosmictron is single-binary and scales well vertically. A 32-core, 128 GB instance handles very large workloads.
2. Module sharding — deploy multiple Cosmictron instances, each owning a shard of sessions by ID prefix, with a lightweight routing layer (nginx, Cloudflare) in front.
3. Read replicas — serve read-only PgWire queries from a snapshot-based replica (manual setup today).