Hosting and cloud

What this shelf covers

Places to run an application: managed platforms that deploy from a repository, container services, virtual machines, and full cloud providers with several hundred services attached.

The right answer depends less on scale than on how much operational work you want to own. Every layer of control you take adds a duty that arrives at an inconvenient hour.

Databases have their own shelf in databases. Watching what happens once it is live belongs in monitoring.

Four levels of control

Managed platforms take code and run it. Deployments come from a repository, scaling is a setting, and most infrastructure decisions are made for you. They are the fastest way to production and the least flexible when something unusual is required.

Container services sit a step lower, running images you build with orchestration you configure. More work, more control, and a portable artifact that moves between providers.

Virtual machines give you an operating system and everything that comes with it, including patching, security and backups.

Full cloud providers offer all three plus a catalogue of managed services, at the price of complexity and a bill that requires attention.

The questions that decide performance

• Region choice, close to your users and to your database.
• Cold starts, on anything that scales to zero between requests.
• Persistent connections, which some serverless models handle badly.
• Build and deploy time, since a ten-minute deploy changes how often people ship.
• Static delivery, cached near visitors rather than served from the origin every time.

Distance between the application and its data is the most common self-inflicted problem. Fifty queries a page across a region boundary produces a slow site that no amount of front-end work can rescue.

Reliability, and what happens on a bad day

Uptime promises are worth reading rather than trusting. Most cover the provider’s own control plane and compensate with service credits, which is not what an outage costs you.

Ask about failure domains: whether your instances sit in one zone or several, what a redeploy does to live traffic, and whether rollback is a button or a rebuild. Ask how logs are retained and whether you can reach them when the dashboard is unavailable.

Then ask about the recovery you control. A provider outage is survivable if your data has a copy elsewhere and your deployment is reproducible. Neither is true by default.

Cost, and the lines that grow quietly

Compute is the number on the pricing page. The invoice is usually made of five things.

Egress is charged per gigabyte for data leaving the provider, and it accumulates fastest in video, downloads and traffic between regions. Storage grows and rarely shrinks. Managed services carry their own meters. Build minutes count on platforms that rebuild on every push. Support plans are extra almost everywhere above the basic tier.

Development and staging environments double much of this unless they are sized down or stopped overnight. Ask what the total looks like at three times current traffic, since that is the figure you will actually meet. The habit of burying such meters is covered in what pricing pages hide.

Portability, and how to keep it

Lock-in in this category is rarely dramatic. It accumulates one convenient service at a time.

Containers and standard runtimes travel. Proprietary functions, queues, identity services and databases do not, and each adoption raises the cost of leaving. That is a legitimate trade when the service saves real work, and a poor one when it saves an afternoon.

Keep infrastructure defined as files rather than as console clicks, and stand the whole thing up from scratch once. What you learn from that exercise is worth more than any comparison table.

Matching the platform to the application

A standard web application with a database is well served by a managed platform, and the time saved is worth the premium for almost any team without dedicated operations staff.

Background processing, queues and scheduled work change the picture, since serverless models handle long-running tasks badly and some platforms have no place to put them at all.

Anything with unusual requirements, whether that is specific hardware, sustained high traffic or strict data residency, tends to push towards raw cloud or dedicated machines.

Static sites belong at the cheap end and often cost nothing to serve properly.

Decide from the shape of the workload rather than from the size of the company. Small teams run demanding applications, and large ones frequently run simple ones.