Beyond Manual Caching: How Readyset Offloads MySQL and Postgres Workloads

Readyset offers two caching modes, deep and shallow, each with distinct tradeoffs and internals. This talk explores the underlying architecture of both: how cache hits are served, how cache misses are handled, and how the system maintains correctness and efficiency.

Deep cache maintains materialized views in a streaming dataflow graph. On a hit, Readyset serves results from that precomputed state with sub millisecond latency. For deep cache, we will see how Readyset executes the query inside its dataflow engine while preserving upstream SQL semantics, correctness, and incremental maintenance. Unlike a classic database executor that chooses execution strategies at query time, deep cache fixes the execution path at materialization time (CREATE CACHE). That plan is used both to maintain the view incrementally as base data changes and to run the query on a miss, which constrains how relational operators are designed and optimized.

Shallow cache provides TTL based result caching without a materialized view. On a hit, Readyset serves the stored result set from memory. Because it bypasses the dataflow graph, shallow cache supports a broader set of queries. On a miss, the query executes against the upstream database; results are keyed by parameterized query template and parameters, with coalescing to prevent thundering herds and configurable refresh behavior. We outline the mechanics of parameterization, TTL, and refresh logic.

The talk concludes with recent improvements in execution internals and the roadmap for optimizing both in engine execution and upstream query paths.

Audience: Engineers interested in database caching, query execution, streaming dataflow, and low latency query serving.