Linux/Unix
Product Overview
Spot instances
Spot instances are just spare EC2 instances, made available at a steep discount of up to 90% lower than the usual on-demand cost of EC2 instances.
The only difference is that they can be claimed back with at least 2 minutes (typically 5-15min) notice when EC2 may need the capacity for other on-demand customers.
In average only about 5% of the Spot instances get interrupted in their first month of usage, and these interruptions can usually be handled automatically.
Would Spot be a good fit for my application?
Many cloud-native workloads are already resilient to Spot interruptions, as they satisfy a few key criteria:
- redundancy: running on multiple instances from an AutoScaling group so that they can sustain the failure of a single instance without any visible impact
- stateless operation, or if stateful and their state can be checkpointed externally or re-created quickly in case of interruptions
- start quickly on new instances, preferably in just a few minutes
- processing happens quickly so that new requests can be migrated to new instances and old requests can complete within the termination notice period
- support running on multiple similar instance types, across generations and families
Some classic examples include:
- Applications running in AutoScaling groups, especially if they scale in/out often and safely handle instance churn
- Stateless web servers
- Applications running in Docker containers or from baked AMIs, which start quickly
- Applications that process data from SQS queues
What does AutoSpotting do?
AutoSpotting migrates existing AutoScaling groups to Spot instances.
It's designed towards reducing the adoption effort as much as possible by not requiring launch template/configuration changes, which can be challenging to perform at scale.
All you need to do is to install it from CloudFormation(it takes about 5 minutes) and tag your AutoScaling groups with with the Key="spot-enabled" and Value="true" as seen below.
How does AutoSpotting work?
It runs in your AWS account and is subscribed to certain EC2 events across all the enabled AWS regions.
Once a new on-demand instance has been launched in an enabled AutoScaling group, it performs the following actions:
- launches an identically configured but cheaper Spot instance from the same or a similar instance type
- swaps this new Spot instance with the On-Demand instance
- cleanly terminates the On-Demand instance, usually before it even had the chance to boot
The Spot instances will be configured from the start as part of the group, and the group will continuously monitor them as it would monitor any other instances.
This process requires no launch configuration changes on your AutoScaling groups and can be deployed at any scale throughout your infrastructure, allowing to reduce costs tremendously.
You can also configure AutoSpotting to keep a percentage or number of OnDemand instances in each AutoScaling, which may desirable in some environments, for example if you have Savings Plans coverage for them.
Many customers install AutoSpotting on hundreds of AWS accounts and empower their teams to enable it on a per-group basis by simply tagging their AutoScaling groups as mentioned above, without the need to configure the instance types needed for each group, which can be tedious on large organizations.
Some customers even run it in the so called "opt-out mode", which takes over all the AutoScaling groups unless they are tagged with the same "spot-enabled" tag but set to "false", a very convenient way to quickly migrate to Spot, especially when combined with the option to keep a percentage of On-Demand instances.
How much can it save?
This heavily depends on your configuration, and on how many groups it can be enabled.
The average savings are around 70-75% of the on-demand prices, significantly more than you can achieve with Reserved Instances or Savings plans, and without any time commitments.
How much does it cost?
AutoSpotting costs less than 5% of the savings it generates over on-demand.
You're charged hourly in increments of $0.001 for each $0.02/hour it saved you.
It is free of charge if it saves you less than $0.02/hour and also for the partial amounts up to the next multiple of $0.02/hour.
Monthly this becomes $0.73 for each $14.6 saved by using it, considering 730 hours/month.
For example if it saves you $40 you'll get charged only $1.46 monthly (for the first $29.2 it saved you), and so on.
It has negligible runtime costs and needs no maintenance due to its simple serverless implementation.
For further information please refer to the documentation available at cloudutil.io or reach out to us on Slack, we're more than happy to help you get started.