Write-ups from deployments we have actually run, organised by section. Specific problems, the reasoning behind the approach, and what it cost to get wrong. Client names are withheld under confidentiality.
Page 2 of 6
These two make opposite design choices about the same problem. OpenVPN negotiates almost everything and is correspondingly flexible; WireGuard fixes almost everything and is correspondingly small. Which is right depends on what your network has to tolerate.
Zabbix and Prometheus are both described as monitoring and are built for different jobs. Zabbix is an integrated system that arrives knowing what a server, a switch and a printer are. That difference is what should decide which one a growing business runs.
Logs answer the questions metrics cannot, and they are the first thing an attacker deletes. Loki makes keeping them affordable by indexing labels rather than log text — which is exactly why the labels have to be chosen carefully.
Custom software projects rarely fail on code. They fail on scope that was never bounded, data that was worse than anyone admitted, and a launch that assumed people would change how they work because they were told to.
Stack debates are conducted on technical merit and decided by taste. For a business commissioning software, the properties that matter are how long the platform is supported, who can be hired to work on it, and what happens when the person who built it leaves.
Automating a bad process produces the same bad outcome faster and with less opportunity to notice. The decision that matters is not which tool to use — it is which processes should be automated, which should be simplified first, and which should be eliminated.
Most business applications do not fail under load. They fail under data — a query that was instant against ten thousand rows and unusable against two million. Scaling is mostly a database problem, and mostly solvable without any architectural change.
Most businesses turn on call recording, keep everything forever, and assume that is the safe option. It is the opposite. Retaining recordings creates obligations, and one of them — card data captured in audio — is a standards violation the moment the recording is written.
VICIdial will run a contact centre on hardware that costs a fraction of the commercial alternatives. What it will not do is tolerate being installed and left alone — almost every failure in a VICIdial estate traces to database growth, carrier quality or dialling configuration rather than to the software.
Doubling call volume does not require double the agents, and the reason it does not is the same reason a small centre feels chaotic while a large one feels calm. Staffing does not scale linearly, and planning as though it does is how centres end up simultaneously overstaffed and missing service levels.
Prometheus is the default metrics store for good reasons, and it has a documented limitation businesses run into by using it for the wrong job. Knowing what it is not for is more useful than another installation guide.
Most server monitoring watches CPU, memory and disk, which are the three things least likely to be the actual problem. The signals that predict outages are duller and are usually not collected at all.
We typically respond within 4–8 business hours.