Cloudflare launches K2 serverless event streams in beta

Cloudflare has introduced K2, a serverless event streaming service designed to address a core challenge in distributed systems: the misalignment between event producers and consumers in scale and timing. Traditional Remote Procedure Call (RPC) architectures often struggle when producers generate more data than consumers can process or when downstream services become unavailable, leading to data loss. K2 aims to solve this by acting as a durable buffer that decouples producers from consumers, allowing each to operate at their own pace.

What K2 offers

K2 is a partitioned, durable log service that stores events as an ordered sequence, enabling multiple consumers to read data independently. Built on Cloudflare’s R2 object storage, it leverages the platform’s global infrastructure to provide high durability (11 nines) and horizontal scalability. Unlike traditional systems like Apache Kafka, which rely on local disks and coordination services, K2 offloads replication and consensus to R2, simplifying the application layer and separating compute from storage. This design allows K2 to handle vast quantities of data while maintaining low operational complexity.

The service supports two primary consumption patterns: partitioned reads, where work is divided among multiple consumers, and pub/sub, where each consumer receives all messages. K2 also includes features like batch leasing, acknowledgment (ack), negative acknowledgment (nack), and lease extension to manage processing reliability. However, its reliance on object storage introduces a trade-off: higher produce latency, currently around 1 second at the 99th percentile, due to the need to accumulate writes before committing them to R2.

Use cases and integration

K2 is positioned as a high-scale data movement and retention solution, distinct from Cloudflare’s existing asynchronous primitives like Queues and Basin Pipelines. Queues are optimized for tracking individual work items with features like retries and dead-letter queues, while Basin Pipelines focuses on serverless ingestion and transformation of events into object storage or Iceberg tables. K2, by contrast, is designed for scenarios requiring long-term retention, fan-out consumption, and custom processing, such as analytics, fraud detection, or real-time monitoring.

Developers can create streams via the Cloudflare dashboard, API, or CLI tools like Wrangler. Events can be produced using HTTP APIs or Worker bindings, and consumers can poll subscriptions to retrieve and process data. The service supports flexible consumption models, including partitioned reads for parallel processing or pub/sub for broadcast-style delivery. During the beta period, K2 is available with limits of 10GB storage and 30 MB/s produce throughput per stream, with higher limits available upon request.

Pricing and roadmap

K2 is currently in public beta for users with Workers Paid subscriptions and will not incur charges during this phase. Once billing begins, Cloudflare plans to charge $0.04 per GB for data produced and consumed, and $0.02 per GB per month for retained data. The roadmap includes enhancements such as higher write parallelism (up to multi-GB/s streams), message keys for ordering guarantees, push-based worker consumers, and lower-latency tiers. Cloudflare also intends to add support for Apache Kafka clients, enabling easier migration for users of existing streaming systems.