How to choose a VPS

Start from the constraint

A virtual server has four things you can run out of: memory, processor time, disk and outbound transfer. Only one of them will be your problem at any given moment, and buying more of the other three is the most common way to spend money without making anything faster. So the first question is not “how big a server do I need” but “which of the four is going to run out first, and what happens when it does”.

The failure modes are different enough to recognise. Running out of memory looks like swapping, then a process being killed, then an outage that resolves itself and comes back an hour later. Running out of processor time looks like everything getting uniformly slower under load and returning to normal the moment traffic drops. Running out of disk looks like a database refusing writes at three in the morning. Running out of transfer looks like an email from us, because we do not cut you off.

Memory is usually the answer

For web workloads, memory is the resource that most often decides how the server behaves, because it is the one with no graceful degradation. A busy machine short of CPU is slow; a machine short of memory starts killing things. An application server and a database on the same box each want their own working set, and the database in particular gets dramatically slower the moment its indexes stop fitting in RAM.

• 1 GB — VPS-1 at $9.90. Enough for a static site, a tunnel endpoint, a bot or a cron box. A database and an application server sharing it will swap under load.
• 4 GB — VPS-2 at $14.90. A single site with its own database starts to behave. Still one shared core, so sustained load is not its strength.
• 8 GB — VPS-4 at $23.90. The size we recommend by default, because the web server and the database stop fighting over one core.
• 8 GB and above — BD-8 at $11.90. Logged-in users, carts and search: the requests you cannot cache away are the ones that need this.

A useful sanity check before you buy anything: on the machine you have now, look at how much memory is actually resident versus cached, and what happens to that number at your busiest hour. If you are already swapping, no amount of extra CPU will help.

Cores: how many, and what kind

Core count is the specification everyone compares and the one that most often misleads, because there are three different kinds of core in our catalogue and they are good at different things.

• Standard — shared vCPU on KVM. Bursts are quick because your neighbours are usually idle. An hour pinned at 100% is the thing a shared core cannot promise.
• High frequency — the same shared model on higher-clocked hosts, with roughly twice the NVMe per gigabyte of RAM. HF-4 gives 2 vCPU and 4 GB for $26.90. Worth it when one thread does the important work.
• Dedicated CPU — cores pinned to your machine that no other tenant can take. DC-8 is $49.90 for 2 pinned cores and 8 GB. You are paying for exclusivity, not for the count.

The comparison that catches people out is BD-8 at $11.90 for 4 shared cores against DC-8 at $49.90 for 2 pinned ones. Fewer cores for more money is not a mistake in the price list: shared is faster when your load is spiky, pinned is faster when it is constant. Pick by the shape of the load, not by the number.

Disk: size for what accumulates

People size disk for what they are about to install, which is almost never the problem. What fills a disk is what accumulates: user uploads, logs nobody rotates, database backups written to the same volume they are backing up, and container images from six deploys ago. An application and its dependencies might be two gigabytes; a year of a modestly busy site can be twenty.

Our plans run from 25 GB on VPS-1 to 22352 GB at the top of the range, all of it local NVMe attached to the host with no storage network in between. Resizing up to a larger plan happens in place; resizing down needs a rebuild, because the disk has to shrink. That asymmetry is a good argument for not buying the biggest disk on day one.

Transfer is outbound only

Every plan has a monthly transfer allowance, from 1 TB upward. Three things about it are worth internalising before you compare plans on it: inbound traffic is free, traffic between your own servers inside one region is not counted at all, and going over does not cut you off — we get in touch instead.

For most sites the allowance is irrelevant and the memory figure decides everything. For anything shipping media, downloads or fat API responses, it is the first column to read. If you are genuinely bytes-out heavy, look at a dedicated server instead, where the uplink is 1 Gbit and unmetered — that is a different billing model rather than a bigger number.

Where the machine should be

Put the server near whoever waits for your responses. Usually that is your users; sometimes it is an API you call on every request, in which case the API’s continent beats the visitor’s. We have 48 regions, and within a line every plan costs the same wherever it runs, so this is purely a placement decision — the only footprint caveat is that the Unmetered and Big disk lines live in their own datacenters.

The thing to know in advance: a region cannot be changed on a running server. Moving means a snapshot, a fresh deploy and a DNS cutover, and the new machine gets a new IP address. It is an hour of work rather than a catastrophe, but it is worth five minutes of thought now. The locations page lists all of them.

Sensible starting points

If you want the short version, these are the plans we would suggest for the four situations people describe most often when they email us.

• A personal site, a VPN or a bot — VPS-1, $9.90.
• One WordPress site with its own database — VPS-2, $14.90, and VPS-4 if it has logged-in users or a plugin stack.
• A production application earning money — VPS-4, $23.90.
• Something that runs flat out all day — DC-8 or larger on the Dedicated CPU line, $49.90.

How to be wrong cheaply

The best argument for a virtual server over hardware is that a mistake costs almost nothing. Every plan bills by the hour as well as by the month, so deploying two sizes, running your own benchmark and destroying the loser costs less than lunch. Resizing up happens in place from the panel and keeps the same IP address. Snapshots are free to take, so the risky upgrade you were putting off becomes a decision you can undo.

What we will not tell you

We do not publish IOPS figures for any plan, because we have not measured them under conditions you could reproduce and a supplier’s headline number would not survive contact with your workload. We do not publish a port speed for the virtual platform, because it is not specified to us in a form we would print as a promise. And we do not publish latency figures for any region, because the number you care about depends on your users’ networks rather than ours.

What we do publish is every specification we actually know, the price you will pay at renewal, and an hourly rate that lets you produce the missing numbers yourself in an afternoon. That is a worse marketing page and a better basis for a decision.