Migration risks for edge-cached key-value stores

Consistency and throughput constraints

Workers KV relies on a hybrid push/pull replication model to manage data. Small objects reside in a distributed database, while larger objects move to R2. This architecture reduces p99 read latencies for sub-1KB objects. The service remains eventually consistent across 180+ locations. Changes may take up to 60 seconds to reach all global nodes because of this consistency model. If your application requires high write throughput to a single key, the one write per second limit on Workers KV makes it an unsuitable replacement for tools like Redis or DynamoDB. You know that Redis handles high-frequency updates to a single key without these constraints. During internal testing, Cloudflare engineers discovered a regression in read-your-own-write consistency. They addressed this with an adversarial test framework to catch edge cases through interspersed reads and writes. This regression showed that reads following writes could return stale data from certain locations. The median response time for KV reads sits at 12ms, but cold reads in a new region can spike to 170ms or more. Cloudflare allows up to 20 namespaces with 1 billion keys in each.

Infrastructure stability risks

Cloudflare faced significant stability issues in August 2026. The company logged 13 separate incidents between August 7 and August 14, 2026, which affected R2, Durable Objects, and Workers KV. An R2 storage failure in the ENAM region on August 7 caused write errors for several buckets. One customer reported that 67GB of data remained unrestored in an R2 bucket days after the incident. On August 8, a loss of dark fiber connectivity in Istanbul caused network performance issues. A Spamhaus listing on August 12 caused a major email security disruption. On August 13, customers encountered increased 503 errors for Magic Transit and elevated errors for Workers KV requests. August 14 brought a Durable Objects and Workflows availability drop, along with HTTP 5xx errors across Middle Eastern and Southeast Asian markets including Bangkok, Jakarta, and Dammam. These failures show the risks of relying on a single provider for both compute and storage. Earlier in 2025, a credential rotation error on March 21 caused an R2 outage that blocked 100% of write operations. In November 2025, a database permissions change caused a ClickHouse query error that triggered a proxy panic. This November outage affected core CDN, security services, and Workers KV. Will Cloudflare stabilize its storage layer enough to prevent another August streak?

Performance and pricing comparisons

Performance benchmarks show distinct differences between edge stores. Cloudflare KV performed the worst in Deno’s benchmark because of its heavy caching and eventual consistency model. Cold reads in a new region require fetching data from central stores, causing latencies to rise. In contrast, Upstash Redis provides faster performance for most operations.

Cloudflare KV supports up to 1 billion keys per namespace. Values reach 2MB in size. The service charges $0.50 per GB of storage above the first GB. This price is higher than Google Firestore, which charges $0.18 per GiB. The $5 monthly Workers Paid plan includes 1 GB of storage and 10 million reads. This plan also includes 1 million writes. Developers face a 50ms CPU limit on the Workers Paid plan. This limit is a significant constraint for compute-heavy tasks. The service also allows for up to 100,000 reads per second per key. This makes it suitable for high-traffic read workloads. Cloudflare KV works best for workloads where the same value gets read many times between writes.