Managed hosting for web apps
We host and operate the web apps we build or take over: system and runtime updates, TLS certificates, nightly off-site backups, monitoring and incident handling. It is meant for systems a business depends on, not for a five-page brochure site. Project support starts at 600 DKK a month; the full page lists what it covers.
Learn more →Cloud architecture on a small-business budget
For a small business, architecture is mostly about deciding what not to run. A typical setup from our own fleet: one VPS with nginx in front, the application as a systemd service or a Docker Compose stack listening only on localhost, the database on the same machine, and a search engine or Redis only where the product needs one. We give each service a memory limit so one process cannot starve the rest, and we size those limits from real behaviour: on one server a search index kept rereading its own data from disk until its limit matched the size of the index. As load grows, we split services across machines before reaching for an expensive managed cloud.
Server and cloud migration
Moving a site or an app to another server is routine; doing it without losing mail, certificates or search rankings takes a checklist. We copy, test the new server before DNS points to it, switch DNS and then retire the old copy. That last step matters: after one site move we found the old server still answering for the domain and serving backup copies of the WordPress configuration file. Migration also covers external services that close down. When the public Danish address service DAWA shut, we built danadresse.dk with the same address IDs, so users could switch to it without rewriting their stored keys.
CI/CD pipelines
Every push should run the tests, and a deploy should be one command that can be undone. For bkg.dk, GitHub Actions runs more than 1,000 tests; the deploy script then checks health, key pages and self-checks on the new release and rolls back on its own if they fail. In a rehearsal where we killed the connection halfway through a deploy, the server returned to the previous release by itself in 285 seconds. danadresse.dk deploys from a signed webhook in about a minute. When GitHub's hosted runners are unavailable, we run tests and builds on a self-hosted runner, while deploys go through a script that always starts with a dry run.
Monitoring and alerting people do not mute
Monitoring that wakes you for every blip gets muted within a week. Our fleet dashboard watches disks, load, memory, services and backups on every server, and an alert gate in front of it sends a recovery notice only after an alert was actually sent, holds low-priority night alerts until the morning and marks anything that fires three times a day as unstable. Each product also has a daily routine that runs its tests and user journeys. The rule we design around: a page that quietly degrades is more dangerous than one that returns an error, so we probe the paths that can fail silently, not only the home page.
Backup and recovery that has been tested
A backup you have never restored is a hope, not a backup. Our servers back up sites and databases every night and copy them off-site to Google Drive, including the system layer: configuration, crontabs and mail. A daily audit reports any server with a missing or stale copy, and we rehearse restores; the latest drill for our phone platform came back 16 of 16 checks identical. Testing pays off: for one client we found that the WordPress database dumps contained placeholder strings where every percent sign should be, so a restore would quietly have damaged page styles and product descriptions.
Docker and containers
We use Docker where it earns its place: apps with several services that must run in known versions. bkg.dk runs as a Docker Compose stack behind an nginx edge container; a new release starts next to the old one, traffic moves only when the new container passes its checks, and the three most recent images stay on the host for a one-command rollback. Python dependencies are locked with hashes, so the production image matches the image that was tested. Docker engine updates across our servers are scripted with a rollback package, and we compare listening ports and site responses before and after each update.
Performance optimisation
Speed work starts by measuring the real page on the real server, not a lab score. On lynbro.dk we found the same fonts downloaded seven times under different URLs and a cookie banner acting as the largest element on mobile; after fixing both, LCP on production fell from 1,196 to 624 ms. Loading the anti-spam widget only when a visitor nears the form cut the start of the home page from 1,211 to 515 KB. On oksana.dk and uahub.dk, page caching for anonymous visitors, OPcache and a Redis object cache brought server response from about 0.9 seconds to between 0.12 and 0.21 seconds. Every change comes with before and after numbers.