Kafka-Style Streams on R2? Cloudflare K2 Enters Public Beta

Hi, Shii-chan here! Here's the announcement of Cloudflare K2, event streams running on top of R2!

blog.cloudflare.com

What was announced?

The Cloudflare Blog announced Cloudflare K2, a serverless event streaming service. It is built directly on R2 object storage for high-scale data movement and long-term retention. By decoupling producers and consumers at the edge, it gives you durable, ordered log streams without the operational overhead of traditional broker clusters.

It is in public beta now, and it requires a Workers Paid plan.

Why it matters

In traditional streaming, producers and consumers must align in scale and in time. K2 solves this by letting independent readers consume at their own pace without losing events.

What changes

You get ordered, durable logs without running broker clusters
Compute and storage are separated, so they scale independently
Long-term retention stays cheap

Dive Deep

How it works

K2 is a partitioned, durable log on top of R2. Object storage has no append operation, so edge services accumulate writes in memory and write complete segment files to R2. Ordering uses R2's atomic operations, so no separate coordination service is needed.

Guarantees and performance

Durability: built on R2's "11 9s" durability
Ordering: events stay strictly ordered within a stream
No data loss: events are never dropped once accepted
Reading: supports both work-splitting subscriptions and pub/sub
Produce latency: about 1 second at p99, because of batching and object storage write speeds

What it looks like

Creating a stream looks like this:

cf k2 streams create --name app_events --http-enabled

Sending from a Worker binding looks like this:

const result = await env.EVENTS.send([{
  content: new TextEncoder().encode(JSON.stringify({event: "page_view"})),
  headers: {"content-type": "application/json"}
}]);

Subscriptions split work among consumers, and messages are leased for 5 minutes. Consumers can ack, nack, or extend leases.

Public beta limits and pricing

Storage: up to 10 GB per account
Produce: up to 30 MB/s per stream
Free during the beta. After the beta: $0.04/GB produced, $0.04/GB consumed, and $0.02/GB/month retained

How it differs from other services

Queues: handles individual work items with retries and dead-letter logic. K2 is for batch data movement at scale
Basin Pipelines: specialized in writing to object storage and Iceberg tables. K2 suits custom processing and varied destinations

Roadmap

Upcoming features include multi-GB/s write parallelism, message keys with key-based ordering, push-based Workers, a lower-latency Express tier, and Apache Kafka client compatibility.

Wrap-up

K2 is a serverless event stream on R2, in public beta
Durable and ordered, with about 1 second p99 produce latency
Free during the beta with 10 GB and 30 MB/s limits, and usage-based pricing later
Different use cases from Queues and Basin Pipelines

This is for anyone tired of running Kafka brokers, or who wants cheaper long-term log retention!