Most SQL Server high availability guides start from the assumption that the database must protect itself, because the hardware underneath it will not. That assumption dates from physical servers. When SQL Server runs as a virtual machine on a hypervisor cluster that detects a dead node and restarts the VM elsewhere, part of the job an availability group used to do is already done. This guide separates the failures each layer handles, puts recovery times and data-loss figures side by side, and prices the second SQL Server licence an availability group needs, so you can decide whether you need Always On, infrastructure HA, or both.
Two different things called high availability
Infrastructure HA works below the operating system. On a Managed Cloud Dedicated Server at MassiveGRID the VM runs on a Proxmox cluster with storage replicated three times on Ceph. If the physical node fails, the cluster detects the loss within 60 seconds and starts the same VM, with the same disks, on another node, usually within about two minutes. SQL Server comes up as it would after a power cut: it runs crash recovery, rolls forward committed transactions from the log, rolls back the incomplete ones, and accepts connections. Nothing inside Windows or SQL Server is configured for this; the VM does not know it moved.
Always On availability groups work inside SQL Server. A second instance on a second server holds a copy of the database and receives every transaction log record as it is written. In synchronous-commit mode the primary waits for the secondary to harden the log before it tells the application the transaction is committed. If the primary fails, a Windows Server Failover Cluster (or Pacemaker on Linux) moves the primary role to the secondary and the listener name follows it. Microsoft's own overview is clear that an availability group fails over at the replica level and that failovers are not triggered by database-level problems such as a suspect database or a corrupt transaction log. Secondary replicas are also, in Microsoft's words, not backups.
The two protect against overlapping but different failures, which is why the honest answer to "do I need Always On" depends on which failures keep you awake.
What each layer protects against
| Failure | Hypervisor HA (VM restart) | Availability group (sync secondary) | Backups |
|---|---|---|---|
| Physical node dies (power, CPU, memory, mainboard) | Yes. VM restarts on another node from the replicated disks. | Yes, if the secondary is on a different node or host. | Last resort |
| Storage failure | Yes. Three replicas on Ceph; a lost disk or node does not lose data. | Yes. The secondary has its own copy of the data files. | Last resort |
| Windows crashes or hangs | Partly. HA can restart a VM that stops responding to the guest agent, but a wedged OS that still answers is not a failure to the cluster. | Yes. The cluster health check sees SQL Server stop answering and fails over. | No |
| SQL Server service crashes | No. The VM is healthy; Windows restarts the service if configured. | Yes, with the flexible failover policy. | No |
| Windows Update or SQL cumulative update reboot | No. The reboot is the downtime. | Yes. Patch the secondary, fail over, patch the old primary. | No |
| Database corruption on disk | No. The corruption is replicated faithfully. | Partly. Automatic page repair can fix single bad pages from the other replica; logical corruption is replicated. | Yes |
| Someone runs DELETE without a WHERE | No | No. The delete is committed on the secondary within milliseconds. | Yes, with point-in-time restore |
| Ransomware encrypts the data files | No | No | Yes, if the backups are off the server |
Three rows matter most. Patching downtime and a SQL Server service crash are the failures only an availability group handles. Corruption, human error and ransomware are handled by neither, which is why Microsoft's business continuity guide says the availability features do not replace a tested backup and restore strategy.
Recovery time and data loss, option by option
Recovery time objective (RTO) is how long the database is unavailable. Recovery point objective (RPO) is how much committed data you can lose. The figures below are what each design achieves by its mechanics, not measurements of a particular server.
| Design | RTO on node failure | RPO | Protects against patch downtime | SQL Server edition |
|---|---|---|---|---|
| Single VM on Proxmox HA with Ceph | About two minutes (60-second detection, then VM boot and SQL crash recovery; a long-running transaction extends the rollback) | Zero for committed transactions; the log is on replicated storage | No | Any, including Express and Web |
| Basic availability group, synchronous | Seconds to tens of seconds, plus application reconnect | Zero | Yes, one database at a time | Standard (two replicas, one database per group, no readable secondary) |
| Availability group, Enterprise, synchronous | Seconds, with listener and read-only routing | Zero | Yes, all databases in the group together | Enterprise |
| Availability group, asynchronous (DR site) | Minutes, manual forced failover | Seconds to minutes of committed data | Yes, manually | Standard (basic) or Enterprise |
| Log shipping | Minutes to an hour, manual role change | The log backup interval, 5 to 15 minutes typically | Yes, manually | Web, Standard, Enterprise (not Express) |
| Daily backups only | Hours (restore time) | Up to 24 hours, or minutes with log backups | No | Any |
Notice that the single VM on an HA cluster already has an RPO of zero for hardware failure, because the transaction log lives on storage that survives the node. What it lacks is the two-minute gap and the ability to patch without an outage.
What the second replica costs
An availability group needs a second SQL Server instance, and that is where the decision usually settles. Three costs stack up.
- A second server. The same vCores, RAM and NVMe again, because a synchronous secondary must keep pace with the primary. On a Managed Cloud Dedicated Server, 4 vCores, 16 GB and 256 GB with Windows Server Standard is $138.68 a month at the Windows rate ($24.75 per vCore plus $2.00 per GB of RAM and $0.03 per GB of NVMe); two of them are $277.36 before any SQL Server licence.
- Standard edition at minimum. Basic availability groups are a Standard edition feature; Express has no Always On at all, and the Web edition that ships with our SQL Server hosting at $33.25 per vCore supports log shipping but not availability groups. Standard is bring-your-own-licence on the managed server or quoted; on the Virtual Private Cloud order form it is listed at $120.00 per vCore a month. A primary with 4 vCores is $480.00 a month in licence alone at that rate.
- The passive replica's licence. Microsoft's SQL Server licensing guide grants fail-over rights to licences covered by active Software Assurance or a subscription: a passive secondary for high availability does not need its own licence as long as it serves no reads, reports or backups beyond maintenance. Licences without Software Assurance, and any secondary that is read from, must be licensed in full. Under the Services Provider Use Rights that govern rented SPLA licences the equivalent passive fail-over rights exist, with the same "passive means passive" condition; confirm the current wording with your reseller before designing around it.
So a Basic availability group on Standard edition roughly doubles the server bill and adds a Standard licence for the primary, and the moment you want to run reports from the secondary you pay a second licence and need Enterprise for readable secondaries. For a database behind one website or one line-of-business application, that is often several times the cost of the application itself.
When infrastructure HA is enough
Infrastructure HA on its own is the right answer for most small and medium databases, and it is worth saying so plainly, because it is cheaper than the alternatives by a wide margin.
- The application can tolerate a two-minute restart a few times a year. Node failures are rare events. A web shop, a booking system, an accounting package or an intranet application that shows an error page for two minutes and then carries on has not lost money in a way that justifies a second server.
- You can schedule patching. Windows Update and SQL Server cumulative updates each need a reboot. If a monthly window at 03:00 local time is acceptable, the availability group's main everyday benefit, patching without an outage, has no value to you.
- The edition is Express or Web. Neither supports availability groups, so the choice is between infrastructure HA plus backups and nothing at all.
- Nobody will maintain a cluster. A Windows Server Failover Cluster needs a quorum witness, a listener, DNS entries, matching logins and SQL Agent jobs on both replicas, and someone who knows what to do when the secondary shows "not synchronizing" at 2 a.m. A VM restart needs nobody.
Pair the single VM with backups that leave the server: a daily full, differential or log backups at the interval your RPO demands, and periodic restore tests. On the managed server MassiveGRID takes daily backups with snapshots; add your own native log backups to a second location if you need minutes rather than a day of recovery point.
When you do need an availability group
- Sub-minute recovery is a contractual requirement. If your SLA to your own customers says 30 seconds, a two-minute VM restart breaks it, and only an availability group with a listener gets you there.
- Patching cannot cause an outage. Systems that run around the clock across time zones, payment switches, hospital systems and trading back ends have no 03:00. Rolling patching through an availability group is the standard way to update them.
- You need readable secondaries. Offloading reports or read-only API traffic to a replica is an Enterprise feature that pays for itself when the primary is CPU-bound by reporting. This is a performance reason more than an availability one.
- You need a second site. An asynchronous replica in another city survives a data centre outage. Infrastructure HA protects against a node, not against the building. MassiveGRID has servers in New York, London, Frankfurt and Singapore, so a Frankfurt primary with an asynchronous London secondary is a two-server design, not a two-provider one.
Combining the two, which is what most serious deployments do
The layers are not alternatives. Run both replicas as VMs on the HA cluster and you get the best of each: the cluster restarts a failed replica's VM in about two minutes, so an availability group that has just failed over is back to two healthy replicas within minutes instead of running exposed on one until someone rebuilds a server. The availability group covers patching, service crashes and the two-minute gap.
Two practical rules when you do this. Place the replicas on different physical nodes, because a synchronous pair on the same host gives you the complexity of an availability group with the single point of failure of a single server; on a managed server, ask us to pin the two VMs to different nodes. And keep infrastructure HA on for the secondary too; a VM that is not restarted automatically is a replica that has to be noticed, diagnosed and rebuilt by a person.
If you want the database layer handled for you, our managed SQL Server hosting gives you the single VM on the HA cluster with Windows Server Standard and SQL Server Web included from $86.34 a month, with the operating system patched, backed up and monitored by MassiveGRID. If you need Standard edition for a Basic availability group, or Enterprise for readable secondaries, see the SQL Server Standard hosting page for how the licence is supplied and ask us to quote the pair.
Frequently Asked Questions
Is a hypervisor HA restart the same as an Always On failover?
No. A hypervisor restart boots the same virtual machine on another physical node after the first one fails, which takes about two minutes on a Proxmox cluster with Ceph storage and involves SQL Server crash recovery. An Always On failover moves the primary role to a second, already-running SQL Server instance within seconds. The restart protects against hardware and storage failure; the availability group also protects against SQL Server crashes and patching downtime.
Can I run an availability group on SQL Server Web or Express edition?
No. Availability groups need Standard edition for a Basic availability group (two replicas, one database per group, no readable secondary) or Enterprise for the full feature. Web edition supports log shipping and failover cluster instances but not availability groups, and Express supports neither. The SQL Server Web licence included with MassiveGRID's managed server is therefore paired with infrastructure HA and backups, not with Always On.
Do I have to pay for the SQL Server licence on the passive replica?
Under Microsoft's licensing guide, a licence with active Software Assurance or a subscription includes fail-over rights for one passive secondary used for high availability, provided the secondary serves no queries, reports or backups beyond maintenance. Without Software Assurance, or if the secondary is read from, it needs its own licence. Rented SPLA licences carry equivalent passive fail-over rights under the Services Provider Use Rights; confirm the current terms with your reseller or with us before you design around them.
Does Always On protect me against corruption or an accidental delete?
No. A delete is committed on the secondary within milliseconds, and most corruption is replicated faithfully, although automatic page repair can fix a single unreadable page from the other replica. Only backups with point-in-time restore, or a deliberately delayed log-shipping standby, protect against those failures. Microsoft's business continuity guide states that the availability features do not replace a tested backup and restore strategy.
What does MassiveGRID recommend for a single business database?
For one database behind one application that can tolerate a two-minute restart a few times a year and a monthly patch window, a single Managed Cloud Dedicated Server on the HA cluster with daily backups and your own log backups is the cost-effective design. Add an availability group when an SLA demands sub-minute recovery, when patching cannot cause an outage, or when you need a second city.
SQL Server on an HA cluster, run for you
MassiveGRID's managed SQL Server hosting puts Windows Server Standard and SQL Server Web on reserved vCores in a Proxmox HA cluster with Ceph storage, from $86.34 a month for 2 vCores, 8 GB and 128 GB in New York, London, Frankfurt or Singapore. We patch, back up and monitor the server; the cluster restarts it on another node if the hardware fails. Need Standard edition for an availability group? Ask for a quote.
See SQL Server hostingFurther Reading
- Microsoft Learn: What is an Always On availability group?
- Microsoft Learn: Basic availability groups for a single database (Standard edition limits)
- Microsoft Learn: Business continuity and database recovery in SQL Server
- Microsoft Learn: Editions and supported features of SQL Server 2022 (high availability table)
- Microsoft: How to license SQL Server (fail-over rights with Software Assurance)