Over the last years, the storage device landscape has considerably diversified. For example, fast PCIe SSDs are now common even for consumers, and byte-addressable Persistent Memory (PMem) blurs the lines between storage and memory. The traditional hierarchy of cache, DRAM, disk that database systems have relied on for decades is an outdated model that is increasingly replaced by a "storage jungle" of devices with different trade-offs.
In this talk, I will present two lines of my work that bring order into this jungle, Mosaic and Plush: Mosaic is a storage engine for *read-heavy* (i.e., analytical) database workloads. It employs a cost model to find and exploit the optimal device makeup for a given workload and can even give purchase recommendations. Plush is a PMem-optimized data structure designed for *write-heavy* (i.e., transactional) workloads. It leverages PMem's low write latency to achieve high throughput while still being crash-consistent and thus freeing database systems from having to keep a log.