The EuroVPS blog
Field notes.
From the engine room.
Decisions, investigations and lessons from running production infrastructure. Written by the people who look after it.
The archive
More from the team.
Our price doesn't change when your traffic does.
Egress per gigabyte, support by tier, burst priced only after it happens. Here is what your bill reads like once you own the machine underneath it.
Don't stop at the first fault.
The failure mode this post kills: your site collapses, someone finds a plausible fault, fixes it, declares victory, and the site stays down. On the worst days there are four more faults behind the first. Here is the untangling method, one rule at a time, with the tell that exposes each kind of fault.
A whole network of websites went slow at once, and not one server was busy.
A network of recruitment sites went slow at the same moment, and not one server was busy. Every tier sat idle except the database, which was waiting on one query that scanned a whole table to return a single row.
Never reboot a mystery: what to capture before a restart
A restart clears the symptom and the evidence with it; on a stateful system the five minutes spent capturing state first decide whether the data survives.
Owned vs rented GPU: utilisation, data locality, support
Renting GPU hours suits bursts; a steady AI workload is better served by a machine you own, for utilisation, data locality and jurisdiction.
When self-hosting an open model beats paying per token.
Steady volume past a threshold favours owned hardware; small or bursty volume favours the API. Here is how to run the numbers with your own figures, and why the answer on cost alone is the one that keeps moving.
Residency is not sovereignty. An EU region in a US console does not change who can be ordered to hand over your data.
A US-headquartered provider running a Frankfurt data center still answers to US data demands. The region picker decides where the data sits, not who can be ordered to produce it.
Your host's storage vendor doesn't know you exist.
The "enterprise-grade storage" badge describes a purchase, not an ability. At most providers a storage fault becomes a manufacturer case moving at the vendor's pace, in a queue where your host is one logo among thousands and you are nobody at all. The question that predicts your 3 a.m.: who at this host can read the array, and how many layers sit between you and them.
You can't load-balance a write.
Identical database nodes look like multiplied write capacity, so you round-robin everything and get deadlock errors handed to editors mid-publish, with no locking bug anywhere in the code. The fix is one elected writer with the reads fanned out, and it costs nothing, because writes are seldom the bottleneck.
The DevOps tax: the real cost of self-hosting AI isn't the GPU.
Teams compare GPU prices to three decimal places, then lose weeks of engineering time to driver and CUDA versions. The expensive part was never the card.
Firewall vs WAF: what each one sees and stops
A stateful firewall reads addresses, ports and connection state; a WAF reads the HTTP request itself, and a login form needs both, in series.
Slow website, idle server: finding where the time goes
The database is blamed first because it is where the wait is felt; measuring each tier of one slow request shows where the time is actually spent.
We put an agency's whole staff on secure remote access in one afternoon.
An agency put its whole staff onto secure remote access in one afternoon, every account with a fixed static IP and full-tunnel egress, and every account came up on the first attempt. Not because bulk provisioning is easy, but because the build was recon first, batch second, and verified by reading the state back instead of trusting the writes.
Server or application bottleneck: how to measure it
Time to first byte at each hop, headroom counters on the box and a slow log keyed by blocking frame place a slow site in one layer before anything is changed.
How to find the one line that makes a checkout slow.
A store feels slow, every server is idle, and the only clue is a feeling. Measure the platform under the failing load, skip the cached homepage, capture the slow requests and bucket them by where they block. The biggest bucket is a file and a line.
How to size a server (most people need less than they think)
Your last invoice records what you were charged, not what you used. Three commands on the box you already run give you the real number.
We're opinionated about hardware so you don't have to be.
A long configurator is homework. We made the storage, network, and compute decisions once, on purpose. This is what we picked, and why.
Ten support tickets in ten years is what boring infrastructure looks like.
A national news publisher has been with us ten years and opened about ten tickets, across election nights. Here is how boring infrastructure is engineered, not stumbled into.
Egress fees exist to make leaving look more expensive than staying.
Data goes into a hyperscaler for free and pays to come out. Four lines from your last invoice show whether that meter is really the barrier.
Buy the serious machine, the way we have since 2004.
Since 2004 we have made the same purchasing decision over and over: buy the serious machine. Here is what that decision buys you.
You don't own your network, and that's your real lock-in.
Most hosts rent their backbone and resell it as their own. Ask one question before you sign: who owns the AS your traffic rides?
Most read
The guides readers come back to- 01Top 10 Reasons Your Site Loads Slowly (And How to Fix Them)
- 02Important Linux Log Files You Must Be Monitoring on RHEL 9 and 10
- 03Configuring WordPress Multisite with Nginx and PHP-FPM on RHEL 9 and 10
- 04Archimedes and the compound pulley: how one man moved a ship
- 05Ultimate Guide to Understanding HTTP Status Codes: 300 Series
- 06Ultimate Guide to Understanding HTTP Status Codes: 400 Series
- 07How much RAM, CPU, Storage, and IOPS resources does your VPS really need?
Put the reading to work
Have a workload in mind?
Start with the numbers.
The sizing desks turn your requirements into a proposed configuration and an indicative price range. An engineer can help you review what fits.
