Hosting
When one server stops being enough, and when it does not
The five signals that a single machine has actually run out, the ones that look like a hardware problem but are not, and what changes operationally the day you own the whole box.
Most teams move to a dedicated server either far too late or for no reason at all. The first group fights a machine that has been out of headroom for months. The second group buys metal to solve a problem that a configuration change would have fixed, and then carries the maintenance cost of a physical box forever.
The useful question is not "is my server slow" but "which specific resource ran out, and can I buy more of it separately".
The five signals that actually mean "out of room"
1. Memory pressure, not memory usage. High RAM usage is normal and healthy; the kernel uses free memory for cache. What matters is swap activity (vmstat 1 showing si/so above zero) and the OOM killer appearing in dmesg. If you are swapping steadily or killing processes to stay up, you need more RAM. That may be a bigger VPS, not a dedicated box.
2. CPU steal time. On a shared or virtualised host, %st in top is the percentage of time your virtual CPU wanted to run but the hypervisor gave the physical core to someone else. Consistently above a few percent is a noisy-neighbour problem. This is the classic case where a bigger plan on the same shared platform does not help, because the constraint is contention rather than size.
3. Disk IOPS saturation. iostat -x 1 with a high %util and rising await means requests are queuing at the device. On virtual machines this is frequently a shared-disk problem, and it produces the maddening symptom of a database that is fast at 3am and unusable at 3pm.
4. A single workload that needs the whole machine. Large databases that want to cache the entire dataset in RAM, heavy compilation, video transcoding, a search index build. These have a genuine floor that no smaller box can meet.
5. Sustained high load average with no headroom left to absorb a spike. If a traffic spike reliably takes you from fine to down, you have no burst capacity, and that is a sizing problem rather than a performance one.
The signals that are not hardware
Before you spend on metal, rule these out. Each one looks like a server that is too small:
- No caching. A WordPress or Laravel site hitting the database on every request will saturate a large machine. Add object caching and page caching and watch the load fall.
- Missing indexes. One unindexed query on a growing table turns a fast box into a slow one. This is the single most common cause of "the server got slow" after a year of growth.
- Unbounded logging. A log directory that nobody rotates will eventually fill a disk, and a full disk looks like a thousand unrelated failures at once.
- Backups running during peak. A nightly
mysqldumpat 9pm on a box that also serves evening traffic competes with itself. - A single runaway process. One bad cron job, one memory leak, one infinite loop. Find it before buying hardware to outrun it.
None of these are fixed by a dedicated server. They are fixed by ten minutes of reading logs.
What genuinely improves on metal
Assuming you have ruled out the cheap explanations, these are the differences that are real rather than psychological:
- No noisy neighbours. The cores are yours. Steal time goes away because there is no other tenant.
- Predictable disk. Local disks rather than a shared storage layer, so IOPS stop depending on what another customer is doing.
- Whole-machine memory. No hypervisor reserving a slice, no balloon driver reclaiming pages.
- Hardware choice. ECC memory, specific CPU generations, multiple disks in a RAID you configure.
- A stable baseline to benchmark against. When you tune something, the number moves because of your change and not because the box beside you got busy.
What does not automatically improve: single-threaded performance on an old CPU, network latency to your users, or anything your code does badly.
What changes operationally
This is the part people underestimate. A VPS and a dedicated server have the same interface (SSH) and a different set of responsibilities:
- Hardware failures are now your problem to notice. On a virtual machine, a dead disk is invisible; the host handles it. On metal, you are the one who sees the SMART warnings, and if nobody is watching them you find out when the array fails.
- Firmware and drivers are yours. Kernel upgrades that touch storage or network drivers can brick a remote machine more easily than they can brick a VM.
- Reinstallation is a deliberate act. It usually means a support request and a wait, not a button that rebuilds the box in ninety seconds.
- Monitoring stops being optional. You need something outside the machine telling you it is alive. A monitoring agent on the machine cannot tell you the machine is down.
If you do not have someone who will read alerts and act on them, a dedicated server is a liability rather than an upgrade. That is not a sales pitch for our VPS plans; it is the reason many teams are correctly on virtual machines.
The middle option people forget
Before metal, there is often a vertical step. A VPS with dedicated vCPU and enough RAM frequently solves the contention problem without the operational weight of a physical machine. You get isolation at the hypervisor level, the provider keeps handling hardware, and you keep the ability to rebuild in minutes.
The decision tree is roughly:
- Are you swapping or being OOM-killed? Buy RAM.
- Is steal time high? Move to dedicated vCPU, which is smaller than a whole server.
- Is disk IOPS the limit? Move to local NVMe.
- Do you need a single workload to consume the entire machine, or do you need specific hardware? Then buy metal.
- If you are not sure which of the above applies, find out first. Guessing is how people overbuy.
When to move
Move when you can name the resource that is exhausted and the number it needs to reach. "We need 128 GB of RAM because the dataset is 90 GB and we want it cached" is a reason. "The site feels slow" is not, and buying a server will not fix it.
Move when you have monitoring that will tell you at 3am. Move when a reboot is something you can plan for rather than fear.
And do not move if a caching layer, an index, or a log rotation would have solved it. Those are cheaper than metal, and they are also cheaper than the VPS you are already paying for.
- vps
- dedicated
- capacity planning

