Why dedicated servers beat public cloud for database workloads

Public cloud is excellent at bursting stateless workers. A primary database is the opposite: it is stateful, latency-sensitive and almost always licensed or tuned against a fixed CPU and disk profile. Putting that workload on a noisy multi-tenant volume is why so many teams discover that “managed SQL” costs more than the server it replaced.

What the database actually waits on

Query time is rarely CPU-bound on a well-indexed OLTP system. It waits on commit latency, checkpoint I/O and buffer-cache hit rate. On a shared volume those three numbers move when a neighbour starts a backup or a migration. On dedicated NVMe the same commit is a local PCIe hop.

We run customer PostgreSQL 16 instances on dedicated servers with local NVMe, not a SAN presented as a cloud disk. WAL sync stays in the same machine as the shared buffers. There is no hypervisor credit system and no “burst balance” that silently expires at 18:00.

Isolation, licensing and clocks

Three operational details show up after the first production incident:

• Noisy neighbour. A shared hypervisor can steal cycles during live migration. Dedicated hardware does not migrate underneath you.
• Licensing. Per-core products (Oracle, SQL Server, some GIS stacks) are priced on the socket you can count. A 64-vCPU cloud SKU is often more expensive than a 16-core EPYC you actually own for the month.
• Clocks and NUMA. Checkpoint and autovacuum behaviour is more predictable when the CPU topology does not change between reboots.

None of this requires a custom kernel. It requires not sharing the disk.

When cloud is still the right answer

Use a managed cloud database when the replica set must span three regions by Friday, or when the team cannot take a backup window. Use dedicated when the working set fits on one or two machines and the SLA is “this instance stays on this PCIe bus.”

A common pattern we see: the application tier on public cloud KVM, the primary and a streaming replica on dedicated in the same data centre. Private VLAN between them; no hairpin through the public internet.

How we provision it

Catalogue dedicated plans include NVMe, L3/L4 DDoS and IPv6. Extra IP space, additional disks and remote hands are scoped on the order — there is no public add-on rate card. If the working set does not match a published SKU, request a quote with vCPU, RAM, disk and replication topology.

The short version: pay for the bus the WAL sits on, not for a virtual disk that happens to say NVMe in the brochure.