Spot Instances

EC2 Spot instances run on AWS's spare capacity at 60–90% below on-demand price. In exchange, AWS can reclaim them with two minutes' notice. For batch workloads, training jobs, and anything that can checkpoint progress, Spot is the right default.

Typical savings

Instance	On-Demand	Spot (typical)	Savings
c7i.large	$0.085/hr	$0.026/hr	70%
g5.xlarge	$1.006/hr	$0.302/hr	70%
p3.2xlarge	$3.060/hr	$0.918/hr	70%
p4d.24xlarge	$32.77/hr	$9.83/hr	70%

Launch on Spot

spawn launch --name training --instance-type g5.xlarge --spot --ttl 8h

spore.host will request a Spot instance. If no capacity is available, it fails immediately with a clear error rather than hanging.

Check Spot prices first

truffle spot g5.xlarge --regions us-east-1,us-west-2,eu-west-1 --show-savings

--show-savings adds On-Demand and Savings columns so you can see the real discount per AZ (on-demand rates are pulled live from the AWS Price List API).

Finding the right instance type

Not sure which instance to use? See Finding the Right Instance for a step-by-step walkthrough: truffle find → truffle search → truffle spot → truffle quotas.

Spot prices vary by region and AZ — sometimes significantly. Running in `us-west-2` instead of `us-east-1` can save an additional 20–30% on the already-discounted Spot price.

## Diversify across instance types

If your workload runs on several instance families, specifying multiple types increases the chance of getting capacity:

```sh
spawn launch \
  --spot \
  --instance-type g5.xlarge,g4dn.xlarge,g5.2xlarge \
  --ttl 8h

spore.host picks whichever type has available capacity first, using the lowest current Spot price.

Handling interruptions

When AWS reclaims a Spot instance, spored receives a two-minute warning via the instance metadata service. It immediately:

1. Sends a spot_interrupt notification to Slack/Teams (if configured)
2. Runs your --pre-stop hook (if set) — use this to save a checkpoint
3. Terminates gracefully

Configure a checkpoint save:

spawn launch \
  --spot \
  --instance-type p4d.24xlarge \
  --ttl 12h \
  --pre-stop "python /home/ubuntu/save_checkpoint.py" \
  --slack-workspace T03NE3GTY

The --pre-stop command has up to 90 seconds to complete on Spot instances — AWS provides only a 2-minute interruption notice and spored needs the remainder for DNS cleanup and shutdown. On non-Spot instances the default is 5 minutes. Override with --pre-stop-timeout. For most checkpoint saves, 60–90 seconds is sufficient.

Resuming after interruption

If you save a checkpoint on interruption, relaunch and resume from it:

spawn launch \
  --spot \
  --instance-type p4d.24xlarge \
  --ttl 12h \
  --command "python train.py --resume-from s3://my-bucket/checkpoints/latest/"

When not to use Spot

• Interactive sessions — a mid-session interruption is disruptive; use on-demand
• Jobs that cannot checkpoint and take more than a few hours
• Strict deadlines — Spot capacity isn't guaranteed at launch time

Finding consistently available types

Some instance types have much more stable Spot availability than others. Use truffle to compare:

truffle find "nvidia gpu" --spot --sort-by-price --region us-east-1

Types with high availability tend to be larger families (more AZs, more capacity pool) and slightly older generations.