Database and compute servers.
Same work, same time.
Analytics, builds, batch jobs and databases on hardware whose speed does not depend on who else is awake. The gain is not raw power, it is repeatability.
- port
- 1, 10, 25, 40, 100 Gbps
- cores
- dedicated
- ecc
- every machine
- locations
- AMS · MEP · IAD · PHX
The same work should take the same time.
Databases, analytics, build farms and batch jobs share one need: a run should take the same time as the last one. On shared hardware it does not, because another tenant's disk or CPU burst changes your timing.
A dedicated machine does not make a job faster than its hardware allows. It makes the result repeatable, and that is what lets you size capacity, compare releases and trust a benchmark.
How to choose.
- 01 ─ working set
Measure the data your queries touch most. If it fits in memory, the disks are left with writes and the cold rest.
- 02 ─ limit
Find out what the job waits on: CPU, memory or disk. Buy for the one that limits you.
- 03 ─ disks
Choose NVMe under a database. Commit and checkpoint latency is disk latency.
- 04 ─ memory
Size memory to the working set. ECC comes in every server, so only the amount is yours to pick.
- 01 ─
Tail latency from a neighbour
on a meter or a shared hostA checkpoint or WAL write waiting behind another tenant's IO shows up as one query ten times slower than the rest.
hereSingle tenant: the IO queue is yours alone.
- 02 ─
Runs you cannot compare
on a meter or a shared hostOn shared hardware the same job takes a different time depending on who else is busy.
hereSame work, same time, run after run.
- 03 ─
Silent corruption
on a meter or a shared hostAn uncorrected bit flip can damage data long before anything alerts on it.
hereECC memory in every server.
What to look for in the machine.
Databases wait on IO latency; NVMe is where that wait gets short.
ECC is in every server; what is left to decide is how much.
Builds and batch jobs spread across cores; more of them, shorter runs.
A GPU is available on a built-to-order machine.
The servers that fit this today are marked under the processor on the price list. The shelf changes; what to look for in the machine does not.
Does the hot set fit in memory.
Set how much data there is and how much of it is read often. Each row is a memory size we build with, around the smallest one that holds it.
carries over: memory 256 GB
Questions to settle before ordering.
- 01 ─
Can I rent a machine for one batch job?
Yes. There is no minimum term and setup is $0. Plans renew automatically, so cancel before the renewal date; cancellation takes effect at the end of the current billing period (Terms, sections 6 and 11).
- 02 ─
Can I choose Xeon or EPYC?
On a custom build, yes: the configurator lets you pick Intel Xeon or AMD EPYC, or leave the choice to us.
The rest, for every machine: all questions and the terms.
- No managed database, no replication setup, no tuning service. We hand over hardware with an operating system on it.[no]
- No hypervisor from us. Install Proxmox or ESXi yourself if you want one.[no]
Tell us what it serves.
The workload, the working set, where the users are. We come back with a machine sized to it, a price and a date.
alsounmetered servers · CDN origins · streaming servers · game servers · storage servers · download servers · virtualisation hosts · blockchain node servers