July 2023 Summaries

As a YC startup, the company deliberately ignores Paul Graham's advice to "do things that don't scale" when developing SpiceDB, an open-source foundation for its products. The development of software is expensive, and scalability is required for authorization systems, which was previously lacking until Google published their internal system, Zanzibar. The MVP's authorization doesn't need to scale, but as the company grows, it will be necessary to address complex access models. Scaling isn't exclusive to traffic, but rather a requirement in various dimensions such as feature development, development velocity, geographies, and eventually traffic. Many teams struggle to implement beyond basic authorization due to lack of focus and time, leading to a breaking point where a new authorization system must be built. The company created SpiceDB to share expertise and connect developers with authorization experts, fostering an ecosystem for secure systems that scale with their needs.

We successfully scaled SpiceDB Dedicated to 1 million requests per second with 100 billion relationships stored, while maintaining sub-problem cache hit ratios above 90% and average latency below 5.76ms for CheckPermission and 74.1ms for WriteRelationships. Our test setup used CockroachDB as the data storage backend, leveraging EKS to distribute pods across virtual machines and isolating workloads onto distinct compute pools. We identified areas for improvement in SpiceDB Dedicated, including connection balancing and pruning, quantization window smearing, and static-CPU management in Kubernetes nodes. The results demonstrate the performance capabilities of SpiceDB Dedicated and provide valuable insights into planning and executing large-scale load tests against CockroachDB.