You manage OS, SQL patching and HA.
Managed OS and engine; validate feature restrictions.
Managed databases; compare tiers, pooling and compute.
Once an offering fits, this is what you still choose inside it. Figures are maximums that depend on hardware, size and region — read them as ceilings, not defaults, and confirm against the linked documentation before you commit to a configuration.
Consider when the workload needs the operating system, a specific SQL build, or an instance feature the managed services do not expose.
Consider this before VM size — disk layout and caching decide SQL throughput on IaaS more often than core count does.
Consider how much availability you want to own. Azure supplies the infrastructure; you design, configure and operate the SQL-level topology.
Consider for most instance-scoped workloads where storage latency is budgeted for and a readable secondary is not required.
Consider when General Purpose fits the shape of the workload but the database count, storage ceiling or I/O has become the constraint.
Consider when the workload needs local SSD latency, a readable secondary, or in-memory OLTP.
Consider for most database-scoped workloads, and wherever an idle database should be allowed to pause.
Consider when the database needs local SSD latency, a read scale-out replica, or in-memory OLTP.
Consider for very large or fast-growing databases, and where read replicas need to scale independently of the primary.
Bundles compute, storage and I/O into one unit rather than letting you size them separately. Suited to customers who want a preconfigured bundle and a predictable monthly figure. Available for single databases and, as eDTU, for elastic pools. You can convert between DTU and vCore.