Technical articles, infrastructure news, and hosting tips from the voxa.host team.
Sizing an OpenSearch or Elasticsearch cluster on bare metal: the heap rule, shard size, why three nodes, a worked disk example, which servers fit hot and warm nodes, and the licensing difference.
Leaving VMware for Proxmox VE: which import path to use, what changes inside Linux and Windows guests, how to size a dedicated host with worked numbers, and a nine step cutover checklist.
How to size a ClickHouse server: cores for parallel scans, RAM for the page cache and GROUP BY, NVMe for merges, the compression math, one node or a cluster, and when a VPS is no longer enough.
How Proxmox Backup Server deduplication works, a datastore sizing method with worked numbers, RAIDZ2 or mirrors, prune and sync jobs, and which server and disk option fits.
Twenty four fast Zen 4 cores with 5.33 MB of L3 each: why the EPYC 9254 suits per core licensed databases, CI and VM hosts, when the dual is worth €230 more, and when to pick a 9554 or 9754.
What a Windows RDP VPS is for, the four plans and prices, the two session limit, bring your own licence, and the firewall, NLA and lockout settings that keep RDP safe on a public IP.
The E5 2630L v3 is the low power version of the 2630 v3: same cores and cache, lower clocks. Why the standard part wins on a rented server, what it suits in 2026, and when to step up.
Black Friday falls on 27 November. Tell real customer load from a flood, protect the uncached pages attackers target, and follow a dated four week plan to choose the right DDoS tier.
How DDoS protection is billed (included allowances, flat capacity tiers, per attack fees, clean traffic and per IP pricing), the hidden costs to check, and three worked monthly totals.
A calm, ordered runbook for a server under DDoS attack: what to prepare, what to check in the first 15 minutes and the next hour, what to send support, how to handle ransom demands and what to fix afterwards.
The network absorbs volumetric floods, but APIs fail on expensive endpoints, login floods and retry storms. Which layer stops which threat, the controls to build, and when the 10 Tbps + Layer 7 tier pays off.
A null route sends all traffic for one IP address to nowhere, the attack and your visitors alike. How blackholing works, how to confirm it from outside, and how to design so one dark address does not take the business down.
An HTTP flood can be tiny in Gbps and still exhaust your CPU and database. What your own stack can handle with caching, rate limits and timeouts, and when upstream Layer 7 filtering is worth paying for.
Protection capacity is the volume filtered upstream before traffic reaches your port. How to size it from your exposure, what the included 100 Gbps covers, and when a 500 Gbps to 10 Tbps tier pays for itself.
Every EPYC server we sell supports AVX 512, while our Xeon E5 servers stop at AVX2. Here is which workloads actually gain from the wider vectors, which gain little, and how to check your software uses them.
Steal time is the CPU your VPS wanted but the hypervisor gave to another guest. Where to read it, what counts as high, what looks like it, and when bare metal is the fix.
RAM runs out before CPU on almost every VM host. A three step sizing method for RAM, CPU and disk, with a worked table showing 30 to 240 guests of 2 vCPU and 4 GB on seven servers.
ECC memory corrects a single flipped bit and stops on worse instead of handing your database silent bad data. What it protects, what it misses, how DDR5 on die ECC differs and how to watch EDAC on Linux.
The EPYC 9554 gives you 64 full Zen 4 cores per socket at 3.1 GHz, the EPYC 9754 gives you 128 Zen 4c cores at 2.25 GHz. Same instructions, same 360 W, same L3. Here is how to choose.
A second CPU doubles cores, memory and memory channels, but each socket owns its own RAM. How NUMA works, how to see it in Linux, and when one socket is the better buy.
We rank seven EPYC and Xeon dedicated servers by monthly price per physical core, per thread and per GB of RAM, then work through a build farm, a database primary and a virtualisation host.
Five small machines caching near your users, with unmetered traffic on every plan, is a fixed cost instead of a bill that grows with success. Also when it is the wrong tool.
A 60 GB disk does not mean 60 GB for user files. The real upload budget per plan, the four options compared honestly, and where a virtual machine stops being the container.
Page loads get slow at random moments because one machine is generating reports, resizing images and answering visitors at once. The fix is a second, smaller machine.
The second cheapest machine you should ever buy is a watcher in a different city. What to measure, what to alert on, and the alerts that only train you to ignore them.
Two requests that sound identical: a server you carve up yourself, and somewhere to run five things. One is bare metal with a hypervisor, the other is three small machines.
Roughly half a gigabyte of RAM per billion parameters at four bits. What that means for each plan, why processor inference is bandwidth bound, and when EPYC is the honest answer.
A persistent desktop in the EU with a fixed address costs a few euro a month on Linux. A Windows application needs our Windows RDP plans in Ashburn, US, or bare metal in Europe, and this is where that line sits.
Your own VPN gives you a clean fixed address and accounts you control, and it makes you responsible for patching it. The two decisions that matter, the setup outline and the honest limits.
One server in one building is a single point of failure. The cheapest fix is a second small machine somewhere else. Recommended pairs, real monthly totals and a failover drill that costs pennies.
A database that no longer fits in memory does not slow down gracefully, it falls off a cliff. How to size for the working set, what our plans really give a database, and when to move to bare metal.
Latency matters for some strategies and is irrelevant for others. How to tell which you have, how to test two cities for a few cents, and why a restart safe bot beats any hosting decision.
Hosted build minutes get expensive fast. What a runner actually needs, which plan fits which pipeline, where to put it, and the three rules that make hourly machines save money instead of costing it.
Real prices by tier, what actually drives them, and the costs that are not on the pricing page — bandwidth overage, extra IPv4, backups, panel licences and setup fees.
The whole path on a fresh Linux VPS: install, proxy, TLS, WebSockets, static files, load balancing, rate limits, caching and hardening — plus what Nginx is not the right answer for.
Two different things sold under one word. What a 1 Gbps port really delivers, why "unmetered" and "unlimited" are not synonyms, and the five questions to ask.
Ubuntu LTS and Debian are the two defaults worth arguing between; Alma and Rocky answer the RHEL question. What actually differs — and the one thing that does not.
Scaling means adding capacity without adding fragility. Four decisions, the point where each stops working, and a costed path from one instance to a split architecture.
A migration is safe when you build and verify the new server before any DNS change. The full sequence — TTL, rsync, hosts-file testing, certificates, cutover and the week after.
Virtualisation is decades old. What changed is sub-minute provisioning plus hourly billing — turning servers from assets you own into resources you allocate.
A VPS is a slice of a machine; a dedicated server is the whole thing. The decision is about variance and duty cycle, not raw power — here is where the crossover actually falls.
One physical machine, no hypervisor, no neighbours. What bare metal actually buys you, which workloads justify it, and the three cases where a dedicated server is the wrong purchase.
No minimum term. Intel Xeon builds carry a delivery estimate of under an hour; AMD EPYC builds are made to order. Every configuration has full root, ECC memory and a guaranteed 1 Gbps unmetered port.
Dedicated billed monthly · No setup fee · Full root on every server