DEVOURED
What Five P99 CONF Talks Reveal About Database Trade-offs
Engineering teams at P99 CONF highlight that database performance optimization is a zero-sum game where solving one bottleneck inevitably shifts the constraint elsewhere.
What: Talks from ScyllaDB, AWS, and turbopuffer explain how they trade RAM for read latency, write amplification for space efficiency, and quantization for search scale.
Why it matters: This reinforces the RUM conjecture: every database design is a deliberate trade-off between Read, Update (Write), and Memory (Space) amplification.
Deep dive
- Object storage (S3) trades low cost for 20-30ms cold-read latency compared to local NVMe.
- ScyllaDB's 'Logstor' engine reduces read amplification by sacrificing RAM index size.
- Valkey improves single-thread throughput by offloading command parsing to I/O threads.
- Quantization (e.g., RaBitQ) enables massive vector searches but increases CPU intensity via reranking.
- Custom object storage clients (e.g., TigerBeetle) can outperform standard SDKs by avoiding blocking I/O and unnecessary allocations.
Decoder
- RUM Conjecture: A theoretical framework stating that database access methods must prioritize one of Read overhead, Update overhead, or Memory overhead at the expense of the others.
- Quantization: A technique in vector search that reduces the precision of vectors to save memory and improve search speed, typically at a small cost to accuracy.
Original article
Full article content is not available for inline reading.
Explore all 69 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi