Oracle Partitioning Policy: Hard, Soft, and Why It Decides Your License Count

Oracle's partitioning policy names nine technologies as hard partitioning and points at two more held in separate documents. That is eleven, and VMware is not among them.

This page lists all eleven, quotes the sentence in Oracle's own PDF saying the policy cannot go into a contract, walks the vSphere version ladder that decides how many hosts Oracle claims, and prices every rung at list.

The file is partitioning-070609.pdf. It runs to two pages, it has never been part of your Oracle Master Agreement, and it routinely decides seven and eight figure numbers. Read those two pages before you read anyone's summary of them, this one included.

What does the Oracle partitioning policy actually say?

It states which technologies Oracle accepts as hard partitioning, meaning those that bind your license to a subset of cores. Everything else is soft partitioning and is licensed as if Oracle could run on the entire physical host.

Oracle publishes this in the Oracle Partitioning Policy document. It is guidance Oracle applies, and it sits outside the Oracle Master Agreement you actually signed.

Is the Oracle partitioning policy contractually binding?

No, and you do not have to argue the point, because Oracle printed the answer at the bottom of its own file. The published copy closes with this:

This document is for educational purposes only and provides guidelines regarding Oracle's policies in effect as of February 14, 2022. It may not be incorporated into any contract and does not constitute a contract or a commitment to any specific terms.

Oracle redates that paragraph each time it reissues the file, so check the date on the copy you download. The operative words have not moved: may not be incorporated into any contract. Print the page, highlight the clause, and put it in the audit binder before the first call.

Then understand what it does not do. Oracle does not need the policy to be a contract term, because your ordering document already defines Processor as the processors where the programs are installed and running. The policy is only Oracle's explanation of what installed covers when a virtual machine can move.

The disclaimer strips the policy of contractual status. It leaves Oracle's reading of the term you did sign completely intact. Those are two different fights, and buyers routinely pick the wrong one first.

• Hard partitioning: Oracle binds the count to the capped or pinned cores.
• Soft partitioning: Oracle counts every core the workload can reach.
• Policy status: guidance Oracle applies, disclaimed by Oracle, never signed by you.
• The trap: never accept boilerplate incorporating Oracle's then current policies into an amendment, an ordering document or a ULA certification letter. One signature converts that disclaimer into a term you are bound by.

What is the difference between hard and soft partitioning?

Hard partitioning physically or firmly limits which cores can run Oracle. Soft partitioning uses software that Oracle says could be reconfigured, so Oracle declines to recognize it as a limit.

The detailed database licensing rules sit in the Oracle Database Licensing Information documentation. The processor count always flows from the partitioning treatment first.

Why does Oracle reject soft partitioning?

Oracle argues that software limits can be changed, so it cannot rely on them. The practical effect is that a single Oracle node on a large host can license the whole host, and on a modern cluster that host is one of many.

There is no point disputing the reasoning. It is a commercial position dressed as a technical one, and it has survived twenty years of buyers pointing that out. Spend the energy on the boundary instead.

Which eleven technologies does Oracle actually approve?

Nine are named in the policy text itself, and two more are admitted by separate Oracle documents that the policy points to. Eleven, and nothing else. If your hypervisor is not on this list, Oracle counts every core it can reach, and the only question left is how far reach goes.

Read the parentheses, because they carry the money

Five of the nine qualify only when the partition is capped. An uncapped IBM Micro Partition, or a Solaris Zone with no CPU cap, is soft partitioning even though the product name sits on the approved list.

Reviewers check the cap, not the badge. We have seen an estate lose a hard partitioning argument on an AIX platform where the capping was correct on eleven partitions and had been removed on the twelfth during a performance incident two years earlier.

Two lines in that table cost buyers more than anything else

First, Oracle VM appears on Oracle's own soft partitioning list by default. It becomes hard partitioning only once vCPUs are pinned with the cpus line in vm.cfg and the setting stays put. Buying an Oracle hypervisor does not buy a smaller count.

Second, the newer approvals live outside the policy PDF entirely. On Oracle Linux KVM your answer is not in the partitioning policy at all. It is in a separate Oracle document that names the olvm vmcontrol utility and warns that the virtual machine has to be stopped and started before the pinning takes effect.

None of this is free. Pinning removes live migration, removes automated load balancing, and removes the automatic restart you bought the cluster for. That is the trade: you surrender mobility to stop paying for it.

Make it an architecture decision with a number attached, which is what the worked count further down is for. Our practical walkthrough of implementing Oracle approved hard partitioning covers the configuration side.

What about Hyper V, Nutanix, and everything the policy never mentions?

Absence from the list is not approval. If a technology is not named as hard partitioning, Oracle treats it as soft, and the counting question becomes how far the management domain reaches.

The boundary follows the management domain

• Microsoft Hyper V. Oracle looks at the failover cluster and the management scope above it, not at the virtual machine's vCPU allocation.
• Nutanix AHV. The Prism management scope defines where a virtual machine can be placed. Our Nutanix Oracle licensing guide works through it.
• Red Hat Virtualization, oVirt, Proxmox, Citrix. Same logic, no named approval, and no published Oracle position to quote back.
• Container platforms. A CPU limit on a pod is a soft limit. The host is the unit Oracle counts.

The practical consequence is uncomfortable and worth saying plainly. Moving off VMware to escape Oracle exposure does not, on its own, change the counting rule. It changes who you pay for the hypervisor.

Why is VMware the partitioning flashpoint?

Because Oracle names VMware on its soft partitioning list, then defines the boundary by where a virtual machine could go rather than where it went. On a current estate that is not the cluster.

Since vSphere 5.1 it has been the vCenter Server instance, and since 6.0 every vCenter instance the environment can reach. The server you installed on stopped being the unit of measurement years ago.

How far does Oracle extend the claim?

As far as live migration reaches, and Oracle defines reach by capability rather than by practice. The question a reviewer asks is not where the database ran. It is where it could have run.

On a linked vCenter estate the honest answer is every host in both instances, which is why the version table below settles more money than any conversation about intent.

How do you contain VMware exposure?

1. Isolate Oracle workloads onto a dedicated, separately licensed cluster.
2. Disable live migration paths into non Oracle hosts.
3. Put the Oracle vCenter on its own single sign on domain.
4. Document the architecture, with dates, so the boundary is provable rather than asserted.

Our guide to Oracle licensing in virtualized environments covers the build side of that list in more detail.

How does your vSphere version change the size of Oracle's claim?

Oracle's VMware position is not one position. It moved three times, each move following a VMware release that removed a constraint on where a running virtual machine could go.

The version in your change record is the largest single input into Oracle's opening number, and that record is yours, not Oracle's.

The 5.1 step is the one buyers miss. Before it, live migration needed shared storage, so the defensible technical boundary was the set of hosts attached to that storage.

Enhanced vMotion removed the shared storage requirement and Oracle's claim moved from the cluster to the vCenter Server instance. From 6.0, cross vCenter and long distance vMotion moved it again, to every vCenter instance in reach.

Two vCenter Servers joined in Enhanced Linked Mode are one boundary as far as a reviewer is concerned, because a migration between them is a right click.

Do DRS host affinity rules help?

They help your evidence. They do not change the policy. Oracle does not accept a VM to host affinity rule as hard partitioning and will say so in writing if you ask.

What the rule changes is what the logs can say, and the logs are the only thing in this argument that can actually be tested. Use the required form, and know exactly why:

• Must run on hosts in group. A required rule. vSphere HA will not restart the virtual machine on a host outside the group. If every host in the group is down, the database stays down. That is the point. The rule holds in the failure case, so the host list you licensed is the host list the database ever touched.
• Should run on hosts in group. A preferential rule. DRS honors it when convenient and vSphere HA overrides it at failover to keep the virtual machine up. One HA event puts your database on an unlicensed host and writes a timestamped record saying so. That event history is precisely what a reviewer asks for.

Build it in this order in vCenter. Cluster, Configure, VM and Host Groups, and create one virtual machine group and one host group. Then Cluster, Configure, VM and Host Rules, Add, rule type Virtual Machines to Hosts, specification Must run on hosts in group.

Record the creation date. A rule created the week the audit letter arrived proves nothing about the eighteen months before it.

The vCenter topology lever

Cross vCenter vMotion between vCenter Servers sharing a single sign on domain in Enhanced Linked Mode is a menu item, so standing the Oracle vCenter on its own domain removes the top rung outright.

Be precise about the limit of that move. From vSphere 7.0 Update 1 the client will also migrate between unlinked vCenter Servers, so a separate domain narrows the default blast radius without making the migration impossible. The affinity rule and the network path still have to do their share.

What does each boundary cost? A worked count

Numbers you can substitute your own into. The estate below is composite, assembled from the shapes we see most often. Every price is Oracle list, before discount, from the Oracle Technology Global Price List.

The core factor is 0.5, the value the Processor Core Factor Table gives modern Intel and AMD x86 server chips. The estate: fourteen hosts, each two sockets by twenty four cores, so 48 physical cores per host.

vCenter A holds two clusters, a production cluster with six hosts and a second with four. vCenter B holds a recovery cluster with four hosts and is joined to vCenter A in Enhanced Linked Mode. Oracle Database Enterprise Edition runs on three virtual machines, resident on two hosts inside the six host production cluster.

Cores at each boundary: two hosts is 96 cores; the production cluster is 6 x 48 = 288; all of vCenter A is 10 x 48 = 480; both vCenters is 14 x 48 = 672. Multiply each by the 0.5 core factor, then by the $47,500 per Processor list price for Enterprise Edition.

The bottom row is what an opening letter looks like when the estate is a linked vCenter pair and nobody has drawn a boundary. $15,960,000 against $2,280,000 of real workload, a factor of seven, from 336 Processors instead of 48.

Our median across soft partitioned estates is 3.5 times, so seven is the bad end rather than the typical one. The arithmetic producing it is identical every time.

What containment buys, in dollars

Move Oracle onto a dedicated three host cluster, in its own vCenter, on its own single sign on domain, behind a must run rule. Three hosts is 144 cores. 144 x 0.5 = 72 Processors. 72 x $47,500 = $3,420,000 at list.

• Against the production cluster claim of $6,840,000, that removes $3,420,000, exactly 50 percent.
• Against the whole of vCenter A at $11,400,000, that removes $7,980,000, exactly 70 percent.
• Against the linked pair at $15,960,000, that removes $12,540,000, or 78.6 percent.

That is where the 50 to 70 percent band on this page comes from. It is not a marketing range. It is the two middle rungs of the ladder above, computed from prices Oracle publishes.

The support line is the number that compounds

Oracle support runs at 22 percent of net license fees a year. On $6,840,000 that is $1,504,800 a year. On $3,420,000 it is $752,400 a year.

The gap is $752,400 every year for as long as the contract runs, and Oracle very rarely lets a support base fall. Over a five year horizon that single scoping decision is worth more than the license difference itself.

Containment is not free either. The three host cluster carries 72 Processors against the 48 you would need if Oracle simply counted the two hosts the databases sit on. That is 24 extra Processors, 24 x $47,500 = $1,140,000 at list.

You are paying that to stop arguing. Inside an audit that has already opened, it is usually the cheapest line on the table.

What the Broadcom pricing change did to this calculation

Dedicating hosts to Oracle costs more than it did in 2023, because VMware itself is now sold as subscription bundles with a core minimum per processor. A three host Oracle cluster is a real line item, not a rounding error.

It is still small against $3,420,000 of Oracle license and $752,400 a year of support. Run both numbers in the same model, and read our VMware licensing change impact analysis before you assume the hypervisor cost kills the containment case. In every engagement we have modeled since 2024, it has not.

Why does the boundary decide your options bill too?

Because options and management packs license on the same Processor count the policy produces. The policy sets one number, and then every option you run multiplies by it.

Note the shape of the problem. A team that enabled the Diagnostics Pack once, on one database, inside a soft partitioned cluster, has created a $1,080,000 list exposure without buying anything or filing a change request.

Find the option usage before you fix the boundary

Run Oracle's own usage script, options_packs_usage_statistics.sql from My Oracle Support Doc ID 1317265.1, on every database inside the boundary. It reports option and pack usage in the same shape Oracle's reviewers use.

• Active Data Guard switches on the moment a standby is opened for reading while redo apply runs.
• Database Vault is installed with every recent database and enabled by a DBA action.
• The Data Masking and Subsetting Pack runs from an Enterprise Manager menu that is visible whether or not you bought it.
• The Diagnostics and Tuning packs are reachable by default unless someone deliberately set the management pack access parameter.

Fixing the boundary without fixing the option usage leaves half the exposure in place, and the half you left is the half that grows quietly.

Does the partitioning policy apply in the cloud?

No. A different Oracle policy applies, and it counts virtual CPUs rather than physical cores. That difference is one of the few structural ways to shrink an exposure rather than argue about it.

What the authorized cloud policy says

The document is Licensing Oracle Software in the Cloud Computing Environment, and it names Amazon EC2, Amazon RDS, and Microsoft Azure as authorized cloud environments.

• You count virtual CPUs, not the physical host, so the whole hypervisor argument disappears.
• Where hyperthreading is enabled, two virtual CPUs count as one Processor license.
• The Core Factor Table does not apply in these environments, so there is no 0.5 multiplier to lean on.
• Oracle Cloud Infrastructure has its own metric again, based on OCPUs and the Oracle service definitions.

Read the current copy of that policy on Oracle's cloud licensing page, because the list of named providers has changed before and will change again.

The cloud rules are a policy too

The cloud document carries the same educational purposes disclaimer as the partitioning policy. It is not a contract term either, and a provider that is not named in it is not covered by it.

So a migration to a cloud outside the named list does not inherit the virtual CPU counting rule. Confirm the treatment in writing before you move a workload for licensing reasons rather than technical ones.

How do you defend an Oracle partitioning claim?

With architecture and evidence, not debate. The strongest position is a contained estate Oracle cannot credibly extend, documented in records that predate the audit letter.

What evidence wins?

Configuration that proves Oracle can only run on the hosts you licensed wins, and it is what Oracle License Management Services ultimately measures. Logs, cluster settings, and migration boundaries beat any conversation about policy intent.

Evidence you generated on a schedule beats evidence you generated after the letter. Quarterly exports, dated and filed, are worth more than a perfect snapshot taken last Tuesday.

What to say on the first call

Say less than you think you should, and say it in writing. Three sentences do most of the work, and each one moves the conversation toward a definition and away from a diagram.

1. Scope. We will provide inventory for the environments where the programs are installed and running. Please confirm which contractual definition your count is based on.
2. Basis. If a broader count is proposed, please identify the contract term that supports it, since policy documents state they may not be incorporated into any contract.
3. Sequence. We will validate your data before responding to any commercial proposal, and we will respond in writing.

Nothing there is aggressive and nothing there concedes anything. It moves the argument onto the ground where a definition, a host list, and a log file decide the number. Our Oracle audit defense team runs this sequence for clients every quarter.

Should you reference the policy into your contract?

• Never accept language incorporating Oracle's then current policies. That converts a disclaimed document into a binding term.
• Negotiate written virtualization terms where the exposure is material, naming the cluster or the hosts.
• Get any agreed boundary into the ordering document, not into an email from an account manager who will move roles next year.
• If you cannot get a term, get silence. Silence leaves you with the Processor definition and a host list, which is a better position than an incorporated policy.

Five artifacts that prove where Oracle actually ran

A reviewer will not accept your architecture diagram. They will accept machine generated records, and five of them decide this argument. Pull all five before you answer a findings letter, because you need to know what they say before Oracle does.

1. V$LICENSE.CPU_CORE_COUNT_HIGHWATER. The highest core count the instance has seen since it last started. If your databases have only ever run on 48 core hosts, this column reads 48. If one landed on a 96 core host during a failover, it reads 96 and no diagram argues that away. Read it beside CPU_CORE_COUNT_CURRENT and CPU_SOCKET_COUNT_HIGHWATER in the same view.
2. DBA_CPU_USAGE_STATISTICS. The high water mark in V$LICENSE resets when the instance restarts. This view does not. It writes a row carrying TIMESTAMP, CPU_COUNT, CPU_CORE_COUNT and CPU_SOCKET_COUNT every time the CPU picture changes, and it keeps them. It is the closest thing to a licensing flight recorder Oracle ships, and it is the first query to run on an estate you have just inherited.
3. review_lite.sql. The Oracle collection script, the one Oracle will send you with a short deadline attached. Run it yourself first, on hardware you control, and read the output line by line. Nothing in an audit goes worse than seeing your own collection output for the first time on Oracle's call.
4. The vCenter migration record. The vpxd log on the vCenter Server appliance, plus the Migrate and Relocate task history in the Events view. This is where a should run rule that broke at 3am shows up as a timestamped line. It is also, when the record is clean, the strongest single exhibit a buyer owns.
5. The VM and host rule export. The rule name, the specification, the membership of both DRS groups, and the date the rule was created. Export it, date stamp it, keep the versions. A boundary you can show has held for two years is worth several times a boundary you can show exists today.

Reconcile two numbers before anything else: CPU_CORE_COUNT_HIGHWATER against your host inventory. If they disagree, Oracle has already won that part of the argument, and you want to learn it from your own query rather than from Oracle's spreadsheet.

Our walkthrough of reading Oracle LMS script output covers what each section of the collection actually proves, and which lines Oracle itself treats as defects rather than usage.

3.5x

Median VMware exposure multiple

50 to 70%

Cut from the cluster and vCenter claims

30 to 40

Virtualization engagements 2024 to 2025

Source: Redress Compliance advisory engagement file, 2024 to 2025.

Where the common advice on the Oracle partitioning policy is wrong

The common advice is to lead with the disclaimer. The policy says it may not be incorporated into any contract, so the claim collapses. We disagree, and we have watched that position burn four to six weeks of an audit clock and finish exactly where it started, because Oracle is not asserting the policy as a term at all. Oracle is asserting the Processor definition you signed, the one counting processors where the programs are installed and running, and using the policy only to explain what it thinks installed covers when a virtual machine can move.

A disclaimer does not beat a definition. A host list beats a definition.

So reverse the order of operations. Week one goes on making the boundary physically true and machine provable: pinned or isolated hosts, a required affinity rule with a creation date on it, a single sign on domain that stops at the Oracle estate, and a V$LICENSE reading that matches your host inventory.

Only once the reviewer has nothing left to extend the count with do you put the disclaimer on the table, as the reason to settle at your number rather than defend the company's own guidance in front of a lawyer. Produced early it is noise. Produced late it moves the number.

What should a buyer do next?

1. Map every host where Oracle currently runs or could migrate, and write down the vSphere version beside each one.
2. Identify whether a recognized hard partitioning option is available on the hardware you already own.
3. Query CPU_CORE_COUNT_HIGHWATER in V$LICENSE and the change history in DBA_CPU_USAGE_STATISTICS on every instance.
4. Run the option and pack usage script on every database inside the boundary, so you price the whole exposure and not just the database line.
5. Check whether the estate runs on one single sign on domain, and split the Oracle vCenter onto its own.
6. Convert every should run affinity rule to must run, and date stamp the change.
7. Isolate Oracle onto a dedicated, separately licensed cluster and disable migration paths into non Oracle hosts.
8. Price all four boundaries at list, including options and 22 percent support, and put the table in front of the budget holder.
9. Document the architecture as provable license evidence, then repeat the export quarterly.
10. Negotiate written virtualization terms into the contract, and never sign language incorporating Oracle's then current policies.

Oracle's VMware position is a policy claim, not a contract term. Before you accept a findings letter built on it, have our specialists challenge Oracle's VMware position with audit defense experts.

Frequently asked questions

What is the Oracle partitioning policy?

It is a two page Oracle document naming the virtualization technologies Oracle accepts as limiting your license count. Nine are named in the text and two more in separate Oracle documents. Everything else, VMware included, is soft partitioning and is licensed as the whole host.

Is the Oracle partitioning policy legally binding?

No. The policy closes with Oracle's own sentence that it is for educational purposes only and may not be incorporated into any contract. That removes its status as a term. It does not remove Oracle's reading of the Processor definition in your ordering document, which counts processors where the programs are installed and running.

What is the difference between hard and soft partitioning?

Hard partitioning firmly limits which cores can run Oracle, so licenses bind to those cores. Soft partitioning uses software Oracle says could be changed, so Oracle counts the entire physical host and, on a cluster, every host the workload can reach.

Which technologies does Oracle approve as hard partitioning?

Physical Domains, capped Solaris Zones, IBM LPAR, capped IBM Micro Partitions, capped vPar, nPar, capped Integrity Virtual Machine, capped Secure Resource Partitions and Fujitsu PPAR are named in the policy.

Oracle VM Server for x86 with vCPUs pinned in vm.cfg, and Oracle Linux KVM with cores set by the olvm vmcontrol utility, are approved in separate Oracle documents. That is eleven, and five of the nine qualify only when the partition is capped.

How does Oracle treat VMware?

As soft partitioning, and the boundary moves with your version. Up to ESXi 5.0 Oracle claimed the cluster attached to shared storage. From 5.1, when Enhanced vMotion removed the shared storage requirement, Oracle claimed every host in the vCenter Server instance. From 6.0, with cross vCenter vMotion, Oracle claimed every host in every vCenter instance within reach.

Does a must run affinity rule count as hard partitioning?

No. Oracle does not accept any VM to host affinity rule as hard partitioning. The rule still matters, because a must run rule is enforced even by vSphere HA, so the database stays down rather than restarting on an unlicensed host.

A should run rule is overridden at failover, and that migration is timestamped in vCenter for a reviewer to find.

Does Oracle VM count as hard partitioning?

Not by default. Oracle names Oracle VM on its own soft partitioning list. It becomes hard partitioning only when vCPUs are pinned with the cpus parameter in vm.cfg, which you then verify in the CPU affinity output.

On Oracle Linux KVM the equivalent is the olvm vmcontrol utility, and the virtual machine must be stopped and started for the pinning to take effect.

Do Hyper V and Nutanix count as hard partitioning?

No. Neither is named in the policy, and absence from the list is not approval. Oracle treats them as soft partitioning and counts by how far the management domain reaches, which means moving off VMware does not by itself change the counting rule.

How do you limit Oracle exposure on VMware?

Isolate Oracle workloads onto a dedicated, separately licensed cluster, put that cluster in its own vCenter on its own single sign on domain, disable migration paths into non Oracle hosts, and document the architecture with dates so the boundary is provable.

Can you argue your way out of a soft partitioning claim?

Arguing the policy is not binding rarely lowers the claim on its own. A contained architecture that proves where Oracle can run is a far stronger buyer position, and the disclaimer works best at the settlement stage rather than the opening one.

Does the partitioning policy apply in the cloud?

No. Authorized cloud environments such as Amazon EC2, Amazon RDS and Microsoft Azure fall under Oracle's cloud licensing policy, which converts virtual CPUs into Processor licenses and states that the Core Factor Table does not apply. That is a separate rule set from the on premises partitioning policy, and a provider not named in it is not covered by it.

Do database options license on the same boundary count?

Yes, and that is where the number doubles. Options and management packs are licensed on the Processor count the partitioning treatment produces. At 144 Processors, Partitioning, Active Data Guard or Database Vault is $1,656,000 at list each, and the Diagnostics Pack is $1,080,000.

How large can VMware exposure get?

In our engagements, soft partitioned estates faced claims a median 3.5 times the cores actually running Oracle. On the fourteen host estate priced on this page the worst rung is seven times: 336 Processors against the 48 the workload needs, or $15,960,000 against $2,280,000 at list.

What does the partitioning boundary cost in dollars?

Take a 48 core host and the 0.5 x86 core factor. A six host cluster is 288 cores, 144 Processors, $6,840,000 at the $47,500 list price. A dedicated three host cluster is 144 cores, 72 Processors, $3,420,000. That is 50 percent, and support at 22 percent of net fees falls from $1,504,800 a year to $752,400.