deniz.in

Markets

Weather

Loading weather

· via Hacker News – Front Page (hnrss.org)

Rivet open-sources S3-backed actor orchestrator with sub-5ms p99 cold starts

YC-backed Rivet has open-sourced an actor-model orchestrator for AI agents that uses S3 as its only infrastructure dependency, claiming 4.99 ms p99 cold starts and 72 KB of memory per actor.

Rivet open-sources S3-backed actor orchestrator with sub-5ms p99 cold starts

Rivet open-sources an S3-backed orchestrator for agents

Rivet, a Y Combinator Winter 2023 company, has released its agent orchestrator as open source under Apache 2.0. According to the project's launch post on Hacker News and its GitHub README, Rivet is built on the actor model: each actor is a long-lived process with in-memory state, a private SQLite database, WebSocket support for realtime updates, and its own scheduling. Actors are addressed by key, so developers can create one per agent, session, user, or tenant and run it on ordinary Node.js, Bun, or Rust.

The performance claims

The launch rests on a handful of figures. Rivet reports 4.99 ms p99 cold starts and 72.4 KB of memory per actor, achieved because actors share a single node process rather than each booting a sandbox; the company says one server can hold hundreds of thousands of them. Its comparison table puts a sandbox approach at 1.70 s p99 cold start and 512 MiB of memory per instance, numbers Rivet draws from a third-party cold start benchmark and a documented default allocation.

Each actor's SQLite database is persisted to S3. The README claims 1.07 ms p99 durable writes and 1.36 ms p99 reads from that setup, at roughly 10.4 times less than replicated SSD storage, a comparison based on S3 Standard list prices against three Amazon EBS gp3 copies. These are the company's own measurements, produced with Rivet 3.0.3 in its published benchmark against Cloudflare Durable Objects; the repository includes the raw samples and scripts for reproducing the results. Idle actors hibernate, cost nothing while asleep, and resume with their state intact on the next request.

S3 as the only dependency

Rivet requires no Postgres, FoundationDB, or NATS. The README states that an object storage bucket is the sole dependency: there is no separate database or message bus to operate, and control plane instances locate each other through the bucket. Besides Amazon S3, the storage layer works with Google Cloud Storage, Azure Blob Storage, and S3-compatible stores such as MinIO, Cloudflare R2, and Hetzner Object Storage. If a node fails, its actors migrate elsewhere carrying their state with them.

Aiming at Durable Objects

Rivet positions itself as a self-hostable, open-source alternative to Cloudflare Durable Objects and says it lifts several of that product's constraints: no 128 MB memory cap, storage beyond the 10 GB SQLite limit, no random eviction, and standard runtimes rather than a custom V8 build. On top of the actor core it adds durable cron and scheduling, event broadcasting, actor-to-actor calls across a cluster, OpenTelemetry traces, typed clients with React hooks, and Vitest support.

For code that an agent writes, Rivet ships agentOS, which executes generated code with granular controls over network access, filesystem access, and resource limits. A Dynamic Apps preview goes further, deploying a generated backend per user the moment it is created, scaling to zero when unused.

Deployment options

The same control plane runs locally and in production. Rivet can be self-hosted on Docker Compose, Kubernetes, VMs, and bare metal, with guides for Hetzner, Railway, Render, and AWS ECS. Rivet Cloud offers fully managed dedicated clusters on AWS or GCP in any region, and a bring-your-own-cloud option runs the control plane inside a customer's VPC on AWS, GCP, or Azure, including air-gapped environments.

Why it matters

Agents are stateful workloads: they accumulate memory, conversation history, and session context that must survive restarts. Giving each one a VM or container makes that state expensive, since idle agents still occupy memory and cold starts still take seconds. Rivet's bet is that object storage, the cheapest durable tier in the cloud, can back per-agent state at single-digit-millisecond latency while hibernation removes idle cost, which would change the economics of running fleets of coding agents or per-user generated backends. The open question is independence: the headline numbers are vendor benchmarks, though the methodology and reproduction scripts are public. The Durable Objects framing also signals where the competition sits, since Cloudflare already sells much of this pattern as a managed service, and Rivet is arguing it can win on cost, density, and control.

  • #open-source
  • #ai-agents
  • #serverless
  • #object-storage
  • #orchestration

Related posts