Readyset offers two caching modes, deep and shallow, each with distinct tradeoffs and internal architectures. This talk explores how both modes work: how cache hits are served, how misses are handled, and how Readyset maintains correctness and efficiency.

Deep cache maintains materialized views using a streaming dataflow graph. On a hit, results are served from a precomputed state with sub-millisecond latency. We will examine how Readyset executes queries inside its dataflow engine while preserving SQL semantics and incremental maintenance. Unlike traditional database executors that choose execution strategies at query time, deep cache fixes the execution plan at materialization time.

Shallow cache provides TTL-based result caching without maintaining a materialized view. On a hit, Readyset returns the cached result set directly from memory. Because it bypasses the dataflow graph, shallow cache supports a wider range of queries. On a miss, the query runs against the upstream database and the result is cached using a parameterized query template and parameters. We also discuss request coalescing, TTL behavior, and refresh logic.

The talk concludes with recent improvements in Readyset’s execution internals and future optimizations for both in-engine execution and upstream query paths.

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