HN Debrief

Celld: Self-hosted, distributed Durable Objects

  • Infrastructure
  • Open Source
  • Developer Tools
  • Databases

Celld is a new Deno project that implements self-hosted Durable Objects. Each object, called a cell, has its own SQLite database, is addressed by name, and is replicated to S3-compatible object storage. The key architectural claim is that nodes coordinate through the bucket rather than through a separate consensus service. That makes the system feel much lighter to deploy than a typical distributed database stack, while still giving developers the familiar “stateful object with local storage” model that made Cloudflare Durable Objects popular.

Treat Celld as a new building block for stateful edge-style apps, not as a drop-in replacement for Cloudflare’s global platform. If you want Durable Object semantics on your own infra and can live with S3 as a hard dependency and control plane, this is worth tracking.

Discussion mood

Strongly positive and curious. People liked seeing the Durable Object model escape Cloudflare, but the enthusiasm was tempered by realism about scope, especially that Celld is not a global edge platform and still depends on S3 as the real coordination layer.

Key insights

  1. 01

    What Celld adds beyond workerd

    The missing piece is cross-node orchestration. Workerd can run Workers code, and Durable Object storage there is just local SQLite, but Celld adds the scheduler and placement logic that let named objects move and run across multiple machines. That turns the idea from a runtime into an actual distributed system you can deploy yourself.

    Do not compare Celld to workerd at the API level alone. If your use case needs object placement, failover, and multi-node execution, Celld is addressing the operational layer that workerd still lacks.

      Attribution:
    • rough-sea #1
    • kentonv #1
  2. 02

    Best fit is single-region ownership

    The credible use case is not replacing Cloudflare’s global network. It is running the same programming model where worldwide placement would be wasted complexity. That narrower target makes the project more compelling, because most teams want stateful event-driven compute under their control, not a planet-scale edge footprint.

    Evaluate Celld for internal platforms, region-bound SaaS workloads, and controlled multi-tenant setups. Skip it if your product promise depends on Cloudflare’s network reach or edge latency story.

      Attribution:
    • benatkin #1
    • wmf #1
  3. 03

    S3 is the hidden control plane

    The “no control plane or consensus” line only holds if you treat S3 as primitive infrastructure rather than part of the distributed system. Commenters were right to call that out. Celld is outsourcing hard coordination problems to object storage and a small set of JSON and replication mechanisms. That is still a smart design choice because it trades bespoke consensus for a dependency most teams already understand and operate against.

    Read Celld as an opinionated stack built on object storage guarantees. Before adopting it, validate the consistency, latency, and failure modes of your S3-compatible backend, because that service now sits on the critical path.

      Attribution:
    • q3k #1
    • jauntywundrkind #1
  4. 04

    The implementation leans on proven pieces

    A commenter highlighted that Celld uses S3 compare-and-swap style coordination and Litestream for SQLite replication. That makes the system less of a moonshot than the marketing might suggest. The novelty is in packaging familiar parts into a Durable Objects deployment model, not in inventing a new storage engine or consensus algorithm.

    This is easier to reason about than many “distributed runtime” launches. If you are considering it seriously, audit the actual behavior of Litestream and your object store first, because those components explain most of the operational risk.

      Attribution:
    • jauntywundrkind #1

Against the grain

  1. 01

    Cloudflare’s network is still the moat

    For workloads that really need broad geographic presence, Celld does not solve the hard part. Durable Objects may live in one datacenter, but Cloudflare still gives you a massive global execution footprint around them. Self-hosting that reach is expensive and operationally ugly, which means Celld only substitutes for a subset of Cloudflare use cases.

    Be explicit about whether you need provider reach or just the programming model. If global ingress and execution are part of the product, Celld will not erase the infrastructure gap.

      Attribution:
    • zuzululu #1 #2
    • vmg12 #1
  2. 02

    Developer onboarding still looks heavy

    A practical complaint was that local experimentation appears to require setting up S3-compatible storage from day one. That is a real adoption tax for a tool aimed at developers who want to prototype quickly. Infrastructure-first ergonomics can keep a promising system from becoming a default choice.

    Watch for a local-only mode or emulator story before betting on broad internal adoption. If your team values fast prototyping, setup friction may matter more than the architecture.

      Attribution:
    • hhthrowaway1230 #1

In plain english

consensus ↩
A class of distributed algorithms that let multiple machines agree on shared state even when some fail or messages arrive late.
Control plane ↩
The management layer of a cloud service where users create, configure, and monitor resources.
Durable Objects ↩
A Cloudflare programming primitive that gives a piece of application state a single authoritative location and compute context.
Litestream ↩
A tool that replicates SQLite databases to object storage for backup and failover.
S3 ↩
Amazon Simple Storage Service, an object storage system whose API is widely copied by other storage providers.
S3-compatible ↩
A storage service that supports the same application interface as Amazon S3, so software built for S3 can often use it without changes.
scheduler ↩
The part of a distributed system that decides where work or stateful objects should run across machines.
single-tenant ↩
A deployment model where one customer or organization gets its own isolated instance of a system rather than sharing it with others.
SQLite ↩
A lightweight embedded relational database stored in a single file and often used for local application state.
workerd ↩
The open-source runtime that powers Cloudflare Workers and can be run on your own infrastructure.

Reference links

Related runtimes and platforms

  • workerd
    Used to compare Celld with Cloudflare’s open source Workers runtime and explain what Celld adds.
  • workerd pull request #6780
    Referenced as work in progress to add the missing scheduler-related capability.
  • Rivet Actors
    Mentioned as another project in the same actor or Durable Object design space.

Background on Deno and Cloudflare

Discussion references