Huong dan toi uu instance spot
Spot Instances Optimization Guide
Spot Instances (AWS), Spot VMs (Azure), and Spot/Preemptible VMs (GCP) are spare cloud capacity offered at 70 to 90 percent off the on-demand gia. The catch is that the cloud nha cung cap can reclaim the instance with as little as 30 seconds of warning when demand for the underlying hardware rises. For tai vuongs that can tolerate interruption — and only for those tai vuongs — spot is the cheapest compute in the public cloud by a wide margin.
What Workloads Belong on Spot
Spot is appropriate for tai vuongs that are stateless, fault-tolerant, and interruptible:
- Batch and HPC jobs — image rendering, video transcoding, Monte Carlo simulations. Checkpoint progress and resume on a replacement instance.
- CI/CD runners — job queues can wait for a replacement instance without missing release deadlines.
- Stateless web tiers — behind a load balancer with autoscaling, an evicted instance is replaced in seconds.
- Big-data processing — Spark, Hadoop, and data-warehouse clusters designed to handle node loss.
- Dev and test environments — non-production tai vuongs where a 2-minute interruption is acceptable.
Spot is not appropriate for stateful databases, single-instance applications with no replica, interactive sessions with users, or batch jobs that cannot be checkpointed.
Provider Comparison
| Provider | Min eviction notice | Max lifetime | Typical discount |
|---|---|---|---|
| AWS Spot Instances | 2 minutes | none | ~70% off on-demand |
| Azure Spot VMs | 30 seconds | none | ~80% off (Linux) |
| GCP Spot VMs | 30 seconds | none | ~60-91% off |
| GCP Preemptible VMs (legacy) | 30 seconds | 24 hours max | ~60-80% off |
GCP Spot VMs (the successor to Preemptible VMs) give the deepest giam gia, and GCP sends a shutdown script 30 seconds before eviction. AWS Spot offers the gentlest notice (2 minutes), giving batch jobs time to checkpoint. Azure falls in between.
Designing for Interruption
The four patterns that make spot reliable:
- Checkpoint and resume — write progress to durable luu tru (S3, Cloud Storage) every few minutes, and restart from the last checkpoint on a replacement instance.
- Distributed work queue — use a queue (SQS, Pub/Sub, Service Bus) to hand out work units. If an instance is evicted, the work unit returns to the queue for another instance to pick up.
- Multi-instance + autoscaler — run a minimum of 2 spot instances behind a load balancer and configure an autoscaler to launch a replacement when one is evicted.
- Spot + on-demand mix — run 70 to 90 percent of capacity on spot and the remainder on on-demand or reserved, so the tai vuong degrades gracefully rather than fully stopping when spot is unavailable.
Spot Fleet and Capacity Rebalancing
AWS Spot Fleet and Azure Spot VM Scale Sets let you request a target capacity across multiple instance types and Availability Zones, which dramatically reduces the chance of all your spot capacity being reclaimed at once. Configure the fleet to span 3 to 5 instance types across 2 to 3 zones, and the probability of total loss drops below 1 percent per month.
GCP does not have a direct Spot Fleet equivalent, but MIG (Managed Instance Group) with multiple instance templates and the proactiveRefresh setting achieves a similar effect.
Cost Tracking
Spot gias fluctuate with demand, so a tai vuong that chi phis $0.029 per hour today might chi phi $0.058 per hour next week. Set billing alerts at 110 percent of the rolling 30-day average, and configure your fleet to fall back to on-demand automatically when the spot gia exceeds the on-demand gia — at that point spot offers no savings and only adds interruption risk.
See our [AWS vs Azure compute so sanh](../../so sanhs/aws-vs-azure-compute/) for the per-instance spot giam gia, and the reserved instances huong dan for the commitment strategy to apply to the on-demand portion of your mixed fleet.