P99 CONF Talks Detail Database Trade-offs in Storage Design and Performance

Database Storage Design Principles

P99 CONF 2026 featured discussions on database storage design, highlighting the RUM conjecture which posits a balance between read amplification, write amplification, and space amplification. Engineers make trade-offs among these factors based on workload and configuration requirements. Improving one aspect often impacts the others.

LSM Trees and Object Storage

Almog Gavra, Co-Founder at Responsive (now Databricks), presented on database design, framing it around read, write, and space amplification. He described Log-Structured Merge (LSM) trees as a mechanism for tuning between write-optimized (log-based) and read-optimized (sorted array) extremes. While originally designed for local disks, LSM trees' immutable files are also suitable for object storage like Amazon S3.

Using object storage introduces trade-offs: objects must be replaced rather than edited, each GET and PUT operation has a cost, and cold reads are slower. For example, an S3 Standard read can take 20–30 milliseconds, compared to 150 microseconds for a local NVMe cache miss. Gavra's SlateDB is an open-source, object-store-native LSM library providing a key-value interface over services like S3, GCS, and R2.

ScyllaDB's Logstor Engine

Michael Litvak and Felipe Cardeneti Mendes from ScyllaDB presented Logstor, an opt-in storage engine designed for read-intensive key-value workloads. Logstor is not a replacement for ScyllaDB's existing LSM engine but an alternative. It uses an in-memory index to store a hash and record location on flash for each key.

This design allows reads to return cached values or locate records with a single disk I/O, bypassing LSM search processes involving SSTables, Bloom Filters, and indexed data structures. Tables using either Logstor or the LSM engine can operate within the same cluster, utilizing ScyllaDB's existing drivers and consistency options. Preliminary tests showed Logstor sustaining 1.1 million reads per second with p99.9 latency under one millisecond.