Learn & Lunch: Write Scaling – Choosing the Right Database 

English Day continues to foster a culture of continuous learning through another insightful Learn & Lunch seminar. 

Introduction 

At Fantastic5 Technology, our weekly Wednesday English Day is dedicated time to sharpen technical communication and deepen engineering knowledge, entirely in English. This week’s Learn & Lunch tackled one of the most impactful decisions behind every on-disk database: how storage engines handle writes, and why getting it wrong costs you at scale.

The session was led by Khiem Nguyen from the Nexus Team, combining theory, benchmarks, and real-world trade-offs over a company-sponsored lunch.

Session Overview

B-Tree vs. LSM-Tree

The session opened by contrasting two dominant storage engine families, each optimized for opposite ends of the read/write trade-off spectrum.

Using random UUIDv4s as primary keys scatters writes across the entire B-Tree index — keeping hundreds of random pages hot in RAM. Switching to UUIDv7 restores sequential patterns and dramatically reduces disk seeking.

LSM-Trees address read amplification through three mechanisms:

• Bloom filters — probabilistic checks that skip files definitely not containing a key
• Block caches — keeping frequently accessed blocks in RAM to avoid repeated disk reads
• Compaction strategies — Size-Tiered (STCS) for write efficiency vs. Leveled (LCS) for read speed

Row store vs. columnar store

Beyond storage engines, the session explored a separate axis: how data is physically arranged on disk.

Lunch and Networking

As the technical deep-dive wrapped up, the team transitioned to the lunch portion. Over a company-sponsored meal, colleagues from different departments continued debating database trade-offs in a relaxed setting – in English, of course. These moments of informal exchange are where ideas move from slides to real implementation discussions.