Australian cloud for business software — built and run in Sydney since 2002

NetLeverage
NetLeverage

Scaling up or scaling out: which parts of a business application can grow which way

Technical article
Hosting

About this article: This article is reviewed against the cited sources. This is source research, not lab validation. Check the stated versions and your own requirements before making changes.

A business application commonly has parts that grow differently. Parts people connect to, such as web front ends and Microsoft remote desktop session hosts, can be spread across more servers when the application supports it. Background workers can be split if the application lets several run at once. In SQL Server, one server accepts changes, so the database grows mainly by being given more processing power, memory and faster storage, with only read-only work able to move elsewhere, and only on some editions. Knowing which part is under strain tells you which kind of upgrade to buy.

Why it matters

When an application slows down, the instinct is to ask for a bigger server. Sometimes that is right. But if the slow part is a session server that could be split across two, a bigger single server costs more and still leaves one point of failure. If the slow part is the database, adding front-end servers will not help at all. Buying the wrong upgrade wastes money and leaves users waiting.

How it works

Two ways to add capacity

Microsoft's architecture guidance defines vertical scaling as "a scaling approach that adds compute capacity to existing resources" and horizontal scaling as "a scaling approach that adds instances of a given type of resource". In everyday terms: scale up means a bigger server; scale out means more servers sharing the work.

The same guidance says vertical scaling suits workloads "that aren't easily divided into smaller parts or when the application architecture doesn't support horizontal scaling", and horizontal scaling suits workloads "that can be divided into smaller parts that run independently".

What makes a part scalable outwards: state

The key question for each part is whether it remembers anything between requests. Microsoft advises: "You should design applications to be stateless where possible. For stateful applications, you should use a state store that's external to your server." A part that keeps nothing locally can run on two, three or ten servers behind a load balancer. A part that keeps a user's session or data on one server needs that user to return to the same server, which limits how freely work can be spread.

Remote desktop session hosts

Microsoft's Remote Desktop Services lets you expand a collection "by adding RD Session Host server virtual machines", and its Connection Broker "can balance the load across the collection's servers when making new connections". If a session disconnects, the broker reconnects the user "to the correct RD Session Host server and their interrupted session". So session hosts scale out, but each user's session lives on one server; other remote desktop products have their own rules.

Classifying the parts of an application

The table is our editorial classification, not taken from a source; your vendor's documentation decides what is supported for your application.

Part State it holds Direction it can grow Notes
Web front end None, or session state that can be moved out Out, if state is external Needs a load balancer and shared session storage if sessions exist
Remote desktop session hosts A user's session while connected Out, by adding session hosts Each user stays on one server for the session; the broker places new sign-ins
Background workers (reports, imports) Work items, if they come from a queue Out, if the application supports several workers Check whether two workers would process the same job
Database (SQL Server) All business data; one primary accepts changes Up for writes; read-only work can move out on some editions See below

The database is the hard part

In SQL Server, one server (the primary) accepts changes. With Always On availability groups, a readable secondary replica allows "read-only access to all its secondary databases", for reports that "can tolerate some data latency". That is an Enterprise edition option. Standard edition's basic availability groups support "two replicas, with one database" and Microsoft lists "No read access on secondary replica", so on Standard a separate reporting copy is the route instead. Our published articles cover the options: Reporting load on SQL Server: readable secondaries, a reporting copy or the same instance and High availability on SQL Server Standard edition.

So unless the business pays for extra database servers and the right edition, database growth means more processing, more memory and faster storage on one server, plus tuning the queries that cause the load.

What to choose and why

  1. Find which part is slow. Measure processor, memory, disk wait and network on each server at the busy time. A slow screen does not tell you which tier is responsible.
  2. If the front end or session servers are saturated, ask the vendor whether more servers are supported. Adding a second can also remove a single point of failure, provided the two run on different hosts and the load balancer or broker in front of them is not itself a single point of failure.
  3. If the database server is saturated, scale it up and look at query and index tuning before anything else.
  4. If reporting load hurts daily work, consider whether a readable copy or a scheduled reporting copy is supported and affordable for your edition.
  5. Record the limit you are approaching so the next upgrade is planned rather than urgent.

A hypothetical example

A 60-person logistics firm runs a dispatch application: web front end, two remote desktop session servers and SQL Server Standard edition. At month-end, users report slow screens. Measurements show the session servers at moderate load and the database server waiting on storage. Adding a third session server would not help. Moving the database to faster storage and adding memory addresses the bottleneck, and moving the month-end report to a scheduled overnight run reduces daytime load. The example is illustrative.

Risks and limits

  • Vendor support decides. An application not designed for several front-end servers may misbehave when given them.
  • Scaling up has a ceiling. One server can only grow so large; past that point the application design must change.
  • Readable copies are not free. They need extra servers, may need a different edition and return slightly stale data.
  • Measure before buying. Upgrading the wrong tier leaves the bottleneck in place.

Sources

Opened 15 November 2026.

Related articles: Sizing a business application server from measurements; Sizing an RDS farm for mixed workloads and one host out of service; the two SQL Server articles linked above.

Related service: Cloud hosting.

Learning notes

  • Scale up adds capacity to one server; scale out adds more servers.
  • Parts that hold no state between requests are the easiest to scale out.
  • Remote desktop session hosts scale out, but each user stays on one server per session.
  • In SQL Server the database scales up for writes; read-only work can move to a readable secondary on Enterprise, or to a reporting copy on Standard.
  • Measure which tier is saturated before choosing an upgrade.