Scaling a Multi-Tenant SaaS Database Without Breaking Performance

Multi-tenant databases work well at low scale, but shared resources magnify inefficiencies as tenant count and traffic grow. The first problems rarely show up as obviously slow queries. Instead, they surface as uneven load, cache contention, and rising tail latency across the system.

What actually breaks at scale:

1. Index for tenant and query shape: Indexing on tenant_id alone isn't enough. Composite indexes must match real access patterns. Misordered indexes often look fine in isolation but fail under concurrent multi-tenant load.
2. Treat query volume as a first-class metric: Many multi-tenant outages come from "fast" queries executed thousands of times per second across tenants. Watching latency alone misses the real source of CPU pressure.
3. Plan for noisy neighbors by default: A few hot tenants can dominate shared resources and quietly degrade performance for everyone else. Isolation and workload-aware routing are far easier to add early than during an incident.
4. Be skeptical of replicas as the long-term fix: Read replicas increase capacity but don't reduce redundant work. At scale, costs grow linearly while lag and operational complexity compound.
5. Eliminate repeated work wherever you can: When the same query shape runs repeatedly with different tenant parameters, recomputation is pure waste. Eliminating redundant execution often delivers larger gains than incremental tuning.

Multi-tenant systems scale best when growth doesn't multiply database work. Readyset removes that repetition at the SQL layer, so teams can scale tenants and traffic without constantly re-architecting their database stack.

🔍 What We're Watching in the Database Ecosystem

AI Zealotry Is Holding Us Back

Matthew Rocklin cautions against uncritical AI enthusiasm. For infrastructure teams, the takeaway is that AI amplifies existing system quality. Strong architecture benefits, weak foundations fail faster.

The Emperor Has No Clothes (in AI)

This piece pushes back on the AI hype cycle, showing how many claims fail under closer inspection. For database and platform engineers, it reinforces the need to evaluate AI-driven systems by production metrics rather than demos or marketing claims.

DBOS: 2025 Year in Review with Mike Stonebraker and Andy Pavlo

DBMS researchers Mike Stonebraker and Andy Pavlo discuss where AI and data systems are heading, cutting through hype with operator and research-grounded predictions. A useful perspective on which AI trends will meaningfully shape database architecture versus those unlikely to survive contact with production.

In Case You Missed It

Join Us for a FOSDEM Happy Hour

We're hosting a happy hour in Brussels after the first day of FOSDEM. If you'll be at FOSDEM or in the area on January 31st, join us for an evening of great conversations with people building database systems. RSVP here to join.