AML LowPriority Compute Still Running Post‑Retirement – Clarification on Spot Conversion.

Question

AML LowPriority Compute Still Running Post‑Retirement – Clarification on Spot Conversion.

Itamar Latnik 40 Microsoft Employee

Hello,
Following the retirement of Azure Batch LowPriority VMs for AML on 31 March 2026, we understand AML Batch‑based clusters must be recreated using Dedicated VMs. However, we are currently migrating our workloads to AKS using Spot nodes as an alternative.

Questions:

Q1 – Runtime Behavior

We are still able to run AML jobs using compute clusters configured with LowPriority VM priority. Can you please confirm whether these clusters are expected to be automatically disabled in the near future, and if so, what the expected timeline is for service enforcement?

Q2 – Spot vs LowPriority Mapping

we are observing successful job execution on clusters that were previously configured as LowPriority. Can you confirm whether AML is internally mapping these retired LowPriority nodes to Spot capacity, or if these jobs are still running on legacy Batch LowPriority infrastructure?

Q3 – Migration Requirement

Given that we are currently able to run AML jobs on existing LowPriority‑configured clusters, is migration to Dedicated AML compute still required, or we must move to AKS‑based Spot node clusters be considered a supported long‑term replacement path for Batch LowPriority AML workloads?

0 comments

Answer accepted by question author

1 additional answer

Your answer

Answer 1

Anshika Varshney 10,315 Microsoft External Staff Moderator

Hi Itamar Latnik,

Thanks for raising this question. This is actually an expected behavior right now, so no need to worry.

Low Priority virtual machines used in Azure Batch were officially retired for Azure Machine Learning support on 31 March 2026. [learn.microsoft.com]

However, even after the retirement date, you may still notice that your existing AML compute clusters which were created earlier using Low Priority configuration are continuing to run jobs successfully.

This happens because Azure is currently performing a system‑initiated migration in the backend where the older Low Priority compute nodes are being converted to Spot based compute nodes automatically. This migration does not interrupt running workloads.

Once this migration is completed:

Existing Low Priority nodes will no longer be available for allocation.

Any new pool creation or scale out request that still uses Low Priority configuration will automatically be fulfilled using Spot instances instead at the Batch layer.

So in simple terms, even though your cluster shows Low Priority configuration, the compute being used underneath may already be backed by Spot capacity.

That said, Low Priority virtual machines are deprecated and will no longer be available for AML workloads moving forward. Spot virtual machines are the recommended modern replacement and are expected to continue as the supported low cost compute option.

Because of this, it is still advised to plan migration of your AML workloads to either:

Dedicated compute clusters in Azure Machine Learning or Spot based compute clusters

so that your workloads remain supported in the long term.

Hope this helps clarify why the existing Low Priority configured clusters are still running jobs post-retirement. Do let me know if you have any further queries.

Thankyou!

Anshika Varshney 10,315 Reputation points Microsoft External Staff Moderator

2026-04-17T06:45:22.78+00:00

Hi Itamar Latnik,

Thank you for sharing the update in the Teams chat. I’m glad to hear that the information was helpful. If you need any further assistance or run into any issues during the cleanup, please don’t hesitate to reach out. I’ll be happy to help.

Since I’ve converted your comment into an answer, could you please take a moment to mark it as Accepted? This will help others in the community who have the same question to find the solution more easily.

Thank you!
Itamar Latnik 40 Reputation points Microsoft Employee

2026-04-26T09:26:45.0433333+00:00

Hi Anshika Varshney,

Thank you for your response, and apologies for the delay—I was out of office.

I have one follow-up question based on your recommendation:

"Because of this, it is still advised to plan migration of your AML workloads to either dedicated compute clusters in Azure Machine Learning or spot-based compute clusters."

Until now, we’ve been requesting new low-priority quota via the portal UI, and there hasn’t been an option to request Spot instances.

Does this mean that going forward, allocations should be done using Spot requests, and that we’re expected to use Kubernetes-based clusters? Thanks,
Itamar
Anshika Varshney 10,315 Reputation points Microsoft External Staff Moderator

2026-04-27T07:39:47.12+00:00

Hi Itamar,

From what you shared, you are correct that earlier it was common to request Low Priority quota directly from the portal. Now the behavior has changed in the backend.

Low Priority virtual machines are no longer offered as a separate option. Azure is moving everything to Spot-based capacity. This means when you try to use Low Priority configuration, Azure may already be using Spot capacity behind the scenes.

So in simple terms, going forward you do not need to look for a separate “Low Priority” or “Spot request” option in the same way as before. Spot is now the standard way Azure provides low-cost, interruptible compute.

About your main question on what you should do next:

You do not have to move to Kubernetes-based clusters unless you already have that requirement for your workloads. Kubernetes with Spot nodes is just one option, not a mandatory path.

You can continue using Azure Machine Learning compute clusters and plan your migration in one of these two directions:

Use dedicated compute clusters if you want stable and predictable capacity Use Spot-based ML compute clusters if you want lower cost and are okay with possible interruptions

Both of these are supported options. The key point is that Low Priority as a separate concept is being phased out and replaced by Spot capacity.

So your next steps can be simple:

Continue your current workloads for now Start planning new clusters using either Dedicated or Spot-based compute Choose based on cost versus stability needs

Hope this gives you a clearer picture. Let me know if you want help.

Thankyou!

Answer 2

Q1 – Runtime behavior

Low-priority VMs in Azure Batch are being retired and automatically migrated to Spot-based compute nodes starting in March 2026. For Batch-managed accounts, existing low-priority compute nodes are converted to Spot-based compute nodes in a system-initiated migration, with no customer action required and no disruption to running workloads. After this migration completes, low-priority compute nodes are no longer allocatable; any pool creation or scale-out request that specifies low-priority nodes is automatically fulfilled with Spot instances instead.

For AML, low-priority capacity is consumed via Azure Machine Learning compute clusters created with tier: low_priority. These rely on the underlying Batch low-priority capability and therefore fall under the same retirement. The context states the Batch low-priority feature retirement and migration window (beginning March 2026) but does not provide an AML-specific shutdown date beyond that. The supported fact is that low-priority nodes will no longer be available after the migration completes and Spot will be used instead at the Batch layer.

Q2 – Spot vs low-priority mapping

After the Batch migration in March 2026:

Any use of low-priority properties in the Azure Batch API or SDK automatically translates into Spot-based compute node allocations.
targetLowPriorityNodes is treated as targetSpotNodes and the pool pricing model changes from low-priority to Spot pricing.

This means that, for workloads still configured to request low-priority nodes (including AML clusters that are backed by Batch and still using a low-priority tier), the underlying Batch service allocates Spot instances rather than legacy low-priority VMs. Jobs continue to run, but the capacity is Spot-based, not legacy low-priority infrastructure.

Q3 – Migration requirement and long-term path

From the Batch side:

Only low-priority compute nodes are being retired; Spot compute nodes remain GA and supported.
After the migration, low-priority nodes are no longer available; all such requests are served by Spot nodes.

From the AML side (based on the provided context):

AML batch deployments can consume low-priority VMs by using compute clusters created with tier: low_priority.
These clusters rely on the underlying low-cost, preemptible capacity model (now Spot after the Batch migration) and automatically requeue failed tasks when capacity is lost.

The context does not state that AML clusters must be recreated as Dedicated-only after March 2026, nor does it declare AKS Spot as the only supported long-term replacement. What is explicitly supported is:

Continued use of Spot-based compute nodes in Batch pools as the long-term replacement for low-priority nodes.
AML batch deployments using low-priority-tier clusters, which conceptually align with this preemptible capacity model.

Therefore, based on the provided information:

Migration is required away from true low-priority Batch nodes, but this is handled automatically by Batch converting low-priority requests into Spot allocations.
Dedicated AML compute is not presented as the only supported path; Spot-based capacity (via Batch Spot nodes) remains supported.
AKS with Spot nodes is not described in the context, so it cannot be confirmed or denied as an officially supported long-term replacement path within this material.

References:

Share via

AML LowPriority Compute Still Running Post‑Retirement – Clarification on Spot Conversion.

1 additional answer

Your answer