Spot instances

1
# cluster.yaml
2
​
3
node_groups:
4
  - name: node-group-1
5
​
6
    # whether to use spot instances for this node group (default: false)
7
    spot: false  # this must be set to true to use spot instances
8
​
9
    spot_config:
10
      # additional instance types with identical or better specs than the primary cluster instance type (defaults to only the primary instance type)
11
      instance_distribution: # [similar_instance_type_1, similar_instance_type_2]
12
​
13
      # minimum number of on demand instances (default: 0)
14
      on_demand_base_capacity: 0
15
​
16
      # percentage of on demand instances to use after the on demand base capacity has been met [0, 100] (default: 50)
17
      # note: setting this to 0 may hinder cluster scale-up when spot instances are not available
18
      on_demand_percentage_above_base_capacity: 0
19
​
20
      # max price for spot instances (default: the on-demand price of the primary instance type)
21
      max_price: # <float>
22
​
23
      # number of spot instance pools across which to allocate spot instances [1, 20] (default: number of instances in instance distribution)
24
      instance_pools: 3
Copied!
Spot instances are not guaranteed to be available. The chances of getting spot instances can be improved by providing instance_distribution, a list of alternative instance types to the primary instance_type you specified. If left blank, Cortex will only include the primary instance type in the instance_distribution. When using instance_distribution, use the instance type with the fewest compute resources as your primary instance_type. Note that the default value for max_price is the on-demand price of the primary instance type, but you may wish to set this to the on-demand price of the most expensive instance type in your instance_distribution.
Spot instances can be mixed with on-demand instances in a single node group by configuring on_demand_base_capacity and on_demand_percentage_above_base_capacity. on_demand_base_capacity enforces the minimum number of nodes that will be fulfilled by on-demand instances as your cluster is scaling up. on_demand_percentage_above_base_capacity defines the percentage of instances that will be on-demand after the base capacity has been fulfilled (the rest being spot instances). instance_pools is the number of pools per availability zone to allocate your instances from. See here for more details.
Even if multiple instances are specified in your instance_distribution, it is still possible that AWS will not be able to provision a spot instance when requested. One possibility is that AWS has exhausted all of the available spot instances of your requested type(s) in your availability zones. Another possibility is that the current price of your requested instance type(s) is higher than your max_price. To mitigate this, you may add a second node group to your cluster configuration which is configured to use on-demand instances as a backup. When doing this, it is important to position the on-demand node group after the spot node group in the node_groups list (since node groups with lower indices have higher priority). See here for docs and examples.
There is a spot instance limit associated with your AWS account for each instance family in each region. You can check your current limit and request an increase here (set the region in the upper right corner to your desired region, type "spot" in the search bar, and click on the quota that matches your instance type). Note that the quota values indicate the number of vCPUs available, not the number of instances; different instances have a different numbers of vCPUs, which can be seen here.
Example spot configuration
Only spot instances
1
node_groups:
2
  - name: cpu-spot
3
    instance_type: m5.large
4
    min_instances: 0
5
    max_instances: 5
6
    spot: true
7
    spot_config:
8
      instance_distribution: [m5a.large, m5d.large, m5n.large, m5ad.large, m5dn.large, m4.large, t3.large, t3a.large, t2.large]
Copied!
3 on-demand base capacity with 0% on-demand above base capacity
1
node_groups:
2
  - name: gpu-spot
3
    instance_type: g4dn.xlarge
4
    min_instances: 0
5
    max_instances: 5
6
    spot: true
7
    spot_config:
8
      on_demand_base_capacity: 3
9
      on_demand_percentage_above_base_capacity: 0
10
​
11
# instance 1-3: on-demand
12
# instance 4-5: spot
Copied!
0 on-demand base capacity with 50% on-demand above base capacity
1
node_groups:
2
  - name: gpu-spot
3
    instance_type: g4dn.xlarge
4
    min_instances: 0
5
    max_instances: 4
6
    spot: true
7
    spot_config:
8
      on_demand_base_capacity: 0
9
      on_demand_percentage_above_base_capacity: 50
10
​
11
# instance 1: on-demand
12
# instance 2: spot
13
# instance 3: on-demand
14
# instance 4: spot
Copied!
Multi-instance
Next - Clusters
Observability
Last modified 2mo ago
Copy link
Contents
Example spot configuration
Only spot instances
3 on-demand base capacity with 0% on-demand above base capacity
0 on-demand base capacity with 50% on-demand above base capacity