Vertical vs Horizontal Scaling

Two ways to add capacity. Vertical scaling (scaling up) means giving a single node more resources: more CPU cores, RAM, faster disks. Horizontal scaling (scaling out) means adding more nodes and spreading load across them via a load balancer or partitioning scheme.

Your service is out of capacity. Vertical scaling is the no-code answer (resize the box, done), and it's the right first move because it requires zero architectural change. But every machine has a ceiling (and the biggest boxes cost wildly more per unit). Horizontal scaling removes the ceiling entirely, at the price of making your app stateless and distribution-aware.

Vertical: stop the instance, pick a bigger instance type, restart, or hot-add resources where supported. Limited by the largest available hardware. Horizontal: run N identical stateless instances behind a load balancer; for stateful systems, shard the data across nodes (see sharding) and replicate for durability. Auto-scaling automates horizontal scaling in response to load.

• Vertical: zero code change, no distribution problems, simplest possible operation
• Horizontal: effectively unlimited capacity, just keep adding nodes
• Horizontal: redundancy comes for free, since one of N nodes dying isn't an outage

• Vertical: hard hardware ceiling, super-linear cost at the top end, and the big box is a single point of failure
• Horizontal: requires stateless app design or sharding, so you've now opted into a distributed system
• Horizontal: needs a load balancer, health checks, and coordination, so operational complexity goes up

• Stock Exchange (Matching Engine) →

  Scales vertically by design: the per-symbol matching engine is single-threaded and in-memory, so a bigger, faster box is the only lever for that hot path

• URL Shortener →

  Stateless read/write services scale horizontally behind a load balancer; Postgres scales vertically first, then with read replicas

• Instagram Feed →

  Feed-serving tier scales out horizontally; the Redis timeline cache scales out via sharding

• Scale vertically first. It buys time with zero architectural change. Only go horizontal when you hit the box ceiling or need redundancy you can't get from one machine.
• A single big box is a single point of failure no matter how reliable the hardware. Vertical scaling alone gives you zero redundancy.
• Horizontal scaling only works if the tier is stateless (or sharded). If instances hold per-user state in memory, you've imported sticky-session problems.
• Cost is super-linear at the top of the vertical curve: the largest instance types cost far more than 2× the mid-tier per unit of compute.