Spot Fleet is an API ideal for workloads that have a start and and end (batch workloads), specially when the workload requires you to have control over which instances to terminate in your own code. You can benefit from integrations with other services such as Load Balancer. However, if you are thinking about using Load Balancers, Auto Scaling Groups are a better option.
Spot Fleet does allow you to diversify across different AZs and networks. Unlike Auto Scaling Group, Spot Fleet, by default, does not take into consideration even re-balancing of instances across AZs (use Auto Scaling Groups instead if AZ-rebalance is key for your workload). Spot Fleet does support types maintain and request. Similar to Auto Scaling groups, the maintain type will preserve the number of instances, and provisioning new healthy instances when it detects that one of the existing instances has become un-healthy.
To support mix instances with different types and purchasing options, Spot Fleet must use Launch Templates. In this exercise, we will re-use the same Launch Template that we created before.
In this part of the workshop we will use an scenario that is common to see when using Spot Fleet: The use of weights and target capacity. This maps well with the concept of “Compute Slots” in batch workloads. Some batch workloads can have more than one worker process running within an instance. The larger the instance, the more worker processes they can run. For example, one worker may consume 1 vCPU: 2GB Ram, which means that on a C5.xlarge (4 vCPUs: 8GB Ram), we could run up to 4 workers. The weight that we will use for the C5.xlarge is 4.
By default, the Target Capacity attribute for Spot Fleet defines the number of instances that the Spot Fleet will procure. When we use Spot Fleet weights, and associate a custom weight with an instance in the override section, the total target capacity is equivalent to the total target of weights that you give to the spot. Going back to the “Compute Slots” or vCPUs analogy, you can instruct Spot Fleet to provide a number of vCPUs or “Compute Slots” from many instances with different sizes.
While in the past weights were only supported on Spot Fleet and EC2 Fleet, now Auto Scaling groups do also support weights and priorities, so now you can consider as well AWS Auto Scaling group for this type of workload.
To start with the hands on part, you first need to create a json file that sets a few relevant Spot parameters; One of the parameters is IamFleetRole. IamFleetRole must be set up to point to the ARN of an IAM role that grants the Spot Fleet the permission to request, launch, terminate, and tag instances on your behalf. For this purpose, you are going to first retrieve the ARN of the Service-Linked role named AWSServiceRoleForEC2SpotFleet. For more information read the spot fleet prerequisites.
If you have never launched a Spot Fleet, the AWSServiceRoleForEC2SpotFleet won’t exist in your account. To create it run the following command in AWS CloudShell: aws iam create-service-linked-role --aws-service-name spotfleet.amazonaws.com
Execute the following command to export the ARN of the Service-Linked role to an environment variable.
export EC2_SPOT_ROLE=$(aws iam get-role --role-name AWSServiceRoleForEC2SpotFleet | jq -r '.Role.Arn')
Copy and paste the following in AWS CloudShell to generate the configuration file:
cat <<EoF > ~/spot-fleet-request-config.json
{
"AllocationStrategy": "capacityOptimized",
"OnDemandAllocationStrategy": "prioritized",
"SpotMaintenanceStrategies": {
"CapacityRebalance": {
"ReplacementStrategy": "launch"
}
},
"LaunchTemplateConfigs": [
{
"LaunchTemplateSpecification": {
"LaunchTemplateId": "${LAUNCH_TEMPLATE_ID}",
"Version": "1"
},
"Overrides":[
{
"InstanceType":"c5.large",
"WeightedCapacity": 2.0,
"Priority": 3.0
},
{
"InstanceType":"m5.large",
"WeightedCapacity": 2.0,
"Priority": 2.0
},
{
"InstanceType":"r5.large",
"WeightedCapacity": 2.0,
"Priority": 1.0
},
{
"InstanceType":"c5.xlarge",
"WeightedCapacity": 4.0,
"Priority": 6.0
},
{
"InstanceType":"m5.xlarge",
"WeightedCapacity": 4.0,
"Priority": 5.0
},
{
"InstanceType":"r5.xlarge",
"WeightedCapacity": 4.0,
"Priority": 4.0
}
]
}
],
"IamFleetRole": "${EC2_SPOT_ROLE}",
"TargetCapacity": 12,
"OnDemandTargetCapacity": 4,
"TerminateInstancesWithExpiration": true,
"Type": "maintain",
"ReplaceUnhealthyInstances": true
}
EoF
Note how the Spot Fleet request file specifies the TargetCapacity and OnDemandTargetCapacity. The split between those two will go into Spot capacity (i.e: to have all on Spot capacity, just set the OnDemandTargetCapacity to 0).
When AllocationStrategy is set to capacityOptimized, Spot Fleet launches instances from optimal Spot pools for the target capacity of instances (in this case weights) that we are launching.
Spot Fleet does support CapacityRebalance as part of the Metadata service. CapacityRebalance is a signal that EC2 emits when a Spot instance is at an elevated risk of being interrupted. Users can intercept this event and decide if they want to launch a replacement by selecting launch as the ReplacementStrategy within the SpotMaintenanceStrategies structure.
By specifying maintain as the request type, Spot Fleet places requests to meet the target capacity over time and automatically replenish any interrupted instances.
If you want to learn more about the other configuration parameters, you can review the documentation here.
Once that you’ve read and familiarised yourself with the configuration, copy and paste this command to submit the Spot Fleet request and export its identifier to an environment variable.
export SPOT_FLEET_REQUEST_ID=$(aws ec2 request-spot-fleet --spot-fleet-request-config file://spot-fleet-request-config.json | jq -r '.SpotFleetRequestId')You have now created a Spot Fleet that uses weights and combines Spot and On-Demand Instances to meet the specified target capacity.
Given the configuration we used above. Try to answer the following questions:
Priority section?1.) How many Spot vs On-Demand Instances have been requested by the Spot Fleet?
This Spot Fleet will maintain a weighted target capacity of 12, as specified in the value for TargetCapacity. Additionally, we are specifying an On-Demand target capacity of 4. Remember that Amazon EC2 calculates the difference between the total capacity and On-Demand capacity in order to launch Spot Instances. This would result in 1 On-Demand Instance being launched and in the case of Spot Instances, given that we are using capacityOptimized allocation strategy, a few things may happen; the most common is that out of the 18 pools, one is selected as optimal (has more capacity, etc). If the selected pool is one of the 9 pools defined as large, 4 Spot Instances will be added. If the pool selected is one of the 9 xlarge, 2 Spot Instances will be created.
2.) For On-Demand Capacity, what does it mean the Priority section?
We did specify prioritized as the allocation strategy for On-Demand Instances. Therefore, the order in which Spot Fleet will attempt to fulfill the On-Demand capacity is defined by the value of Priority in the list of overrides. It will aim first to launch capacity from the c5.xlarge pools with the highest priority. If there are Insufficient Capacity Errors, it will go into the pools with the next higher priority m5.xlarge, and so on.
For Spot, the priority attribute will have no effect in the example. We have specified capacityOptimized as allocation strategy to follow the best practices associated with Spot. This will make the Spot Fleet launch instances from Spot Instance pools with optimal capacity.
3.) How can we check the state of the Spot Fleet created?
While the command below will provide us the description of the Spot Fleet Request
aws ec2 describe-spot-fleet-requests --spot-fleet-request-ids=$SPOT_FLEET_REQUEST_ID
To get the description of the instances and what happened over time to the request we can use:
aws ec2 describe-spot-fleet-instances --spot-fleet-request-id=$SPOT_FLEET_REQUEST_ID
There is also a describe-spot-fleet-request-history that will showcase the steps the fleet took to procure the capacity. You can request it using the following command.
aws ec2 describe-spot-fleet-request-history --spot-fleet-request-id=$SPOT_FLEET_REQUEST_ID --start-time=$(date '+%m-%d-%YT%H:%M:%S' -d '-1 hour')
There is an allocation Strategy named capacityOptimizedPrioritized. What would be the results if you repeat the same exercise with this allocation strategy?
These are some of the features and characteristics that Spot Fleet provides, in addition to the ones covered in this section:
LaunchTemplateConfigs section of the configuration file. By doing so, you can launch a Spot Fleet with instances that vary by instance type, AMI, Availability Zone or subnet.You can read more about Spot Fleet here, in the AWS Spot Fleet documentation.