Notesfoundations🌱 seedling
Scalability
Also: Scalable System · Scaling
Scalability is a system's ability to cope with increased load. It isn't a yes/no property: it only means something once you say which load grows, by how much, and what performance you need to keep.
Why it exists
The most common reason a working system degrades is growth, not bugs: 10× the users, 1,000× the data. Scalability is the discipline of asking, before it happens, "if load grows this way, what are our options?"
How it works
Three steps:
- Describe the load with a few Load Parameters: requests per second, read/write ratio, data size, fan-out, and the shape of skewed distributions.
- Describe performance as a distribution: percentiles for online systems, throughput for batch.
- Add capacity: Vertical Scaling (a bigger machine), Horizontal Scaling (more machines), or a design change such as Fan-Out on Write that moves work off the hot path.
- Queueing explains why performance falls off a cliff near saturation; Little's Law links throughput, latency and concurrency.
Trade-offs
| Gains | Costs |
|---|---|
| headroom for growth without a rewrite | architecture tuned for one load shape is wrong for another |
| options ready before you need them | building for scale you don't have is expensive and slows you down |
In practice
- ChatGPT's PostgreSQL (2026): one primary plus ~50 read replicas served 800M users; shardable, write-heavy workloads are being moved to sharded systems.
- Twitter timelines (2012): the hard load parameter wasn't tweets per second but followers per account.
Pitfalls
- "It scales" with no load parameter attached is a sentiment, not a property.
- An architecture for a million users is usually wrong for a hundred, and the other way round.
- Elasticity (automatic scaling) helps unpredictable load but adds operational surprises; manual scaling is simpler.
Connections
Recall
These cards are also in your review deck, scheduled with spaced repetition.