Rethinking Memory Management for Azure HorizonDB on Disaggregated Storage

Really interesting breakdown of how disaggregated storage changes PostgreSQL’s memory assumptions.

The shift from relying on the OS page cache to treating shared_buffers as the primary cache is especially important. The sections on retiring ring buffers and prefetching directly into shared buffers clearly show how architecture decisions that work well for local storage can become inefficient once remote page servers are introduced.

The prompt-caching analogy is also interesting from an AI systems perspective: both cases show that performance gains often come from restructuring where repeated work happens, rather than simply adding more resources.