Batch Processing
Azure Batch · Spot vs dedicated nodes · HPC
Practice questions for this topic →
Spot vs dedicated nodes
| Factor | Favor Spot | Favor dedicated |
|---|---|---|
| Task recovery | Short, restartable or checkpointed tasks | Interruption loses costly work |
| Completion time | Flexible deadline | Strict deadline or predictable capacity |
| Workload shape | Many independent tasks | Long-running, tightly coupled multi-node MPI jobs |
| Capacity | Pool can run at reduced capacity | Pool must keep its full target capacity |
Facts to know
- Spot has no SLA. Nodes can be preempted at any time, even immediately after creation. Interrupted tasks are requeued automatically; work that wasn't checkpointed is lost, and checkpointing is up to you.
- A preempted Spot VM may be restored, but only during the first 48 hours and with no guarantee.
- You can combine dedicated and Spot nodes: dedicated as the baseline or fallback, Spot for cheaper extra capacity.
- Before setting Spot targets, plan capacity: quota, regional capacity, VM availability and fallback options.
- Low-priority nodes retired on Sep 30, 2025; Spot is the current term and is generally available.
Exam Tips
Tick each tip once you can answer it without looking.
“Large-scale parallel or HPC jobs scheduled across a pool of VMs, no cluster to manage” → Azure Batch.
“Cut Batch cost; tasks are restartable and the deadline is flexible” → Batch pool with Spot nodes. Trap: Spot has no SLA and can be preempted immediately after creation.
“Tightly coupled MPI job with a strict deadline” → Dedicated nodes using RDMA/InfiniBand-capable HPC VM sizes.
“Balance low cost with guaranteed progress” → Mix dedicated baseline nodes + Spot nodes in the pool.
“Low-priority VMs in Batch” → Retired Sep 30, 2025: the answer is now Spot VMs.
“What happens to a task when a Spot node is preempted?” → It's requeued automatically; checkpointing to durable storage is your responsibility.
“Short event-driven background jobs (not HPC)” → Functions or Container Apps jobs. Trap: Don't over-engineer with Batch.
“Preempted Spot node in Batch” → Azure tries to restore the VM only during the first 48 hours, with no guarantee. Trap: Interrupted tasks are requeued; work not checkpointed is lost.
“Before setting Spot node targets” → Plan capacity: quota, regional capacity, VM availability and fallback options. Trap: Spot has no SLA and can be preempted immediately.
“Pool must hold its full capacity or meet a strict deadline” → Dedicated nodes (mix in Spot only for extra cost-savings). Trap: Tightly coupled MPI jobs also favor dedicated nodes.