Dicer is an infrastructure system designed for building sharded services, focusing on low latency and high availability by colocating in-memory state with computation. In traditional stateless models, every request hits a database, which increases costs and latency. Remote caches help but still suffer from network latency, CPU overhead from serialization, and inefficiencies from overreading data. Dicer addresses these issues by allowing applications to use a small library that integrates with a control plane. This system manages shard assignments dynamically, responding to various signals like application health and load, ensuring services remain available and balanced even during disruptions. Databricks employs Dicer in critical workloads, achieving significant performance gains. For instance, migrating the Unity Catalog to a sharded service with Dicer reduced database load by over 10 times. The SQL Query Orchestration Engine saw improved availability and zero downtime during rolling updates, thanks to Dicer’s dynamic sharding. Features like state transfer, which will be available in future releases, allow seamless value transfers between pods during planned restarts, minimizing impact on cache hit rates. Dicer operates by modeling applications around logical keys, assigning these keys to pods through a system of slices that can automatically adjust to maintain balance. The Assigner service generates and distributes assignments, while client libraries (Slicelet and Clerk) facilitate fast lookups and local caching. Dicer also includes advanced features such as zero downtime during restarts, crash detection, load balancing, and fast hot key isolation. The system can replicate hot keys across multiple pods to handle high demand scenarios, making it particularly suitable for read-heavy applications. To deploy Dicer, users can quickly clone the repository and follow instructions in the demo section.