AWS EC2(Spot)

Overview

• Opportunity to reduce EC2 costs by running on unused AWS capacity

Spot pool (capacity pool)

• Set of instances with the same:
  • AZ
  • instance type
  • instance size
  • misc
    • OS
    • network platform (EC2-Classic or EC2-VPC)
• Each pool has its own number of instances available
• AWS continuously monitors it and updates spot price
• Shared between on-demand instances and spot instances

Pricing

• fluctuates (infrequently) based on supply/demand
• you pay actual price (often lower than your bid)
• everyone pays the same price
• partial hours
  
  • not billed when interrupted (i.e. last hour is free)
  • billed when terminated by customer (i.e. last hour is paid)

Interruption

• AWS reclaims the instance back
• Application should be fault tolerant (must be prepared)
• Does not happen very often
• Can hibernate
  • Memory is flushed to disk (just like laptop)
  • Started again when capacity is available
• When
  • (obsolete) bid price < market price
  • Capacity: not enough unused instances
    • on-demand have priority
  • Constraints: e.g. launch group or AZ can no longer be met
• How
  • 2 minute termination notice
  • can be retrieved via metadata 
  • CloudWatch Event

Spot fleet

• Fleet of spot instances managed by AWS
• Spot fleet request
  • Target capacity (how many instances you want)
  • Launch Specifications 
  • ValidFrom
  • ValidUntil
• EC2 tries to maintain target capacity by looking at all specifications (spot pools) and give most economical combination
• Resource-oriented
  • Specify weights for instance types ("capacity units" for whatever you are after: CPU, memory, bandwidth, TPS)
    • WeightedCapacity how many capacity units instance types contributes to the fleet
      • "I am OK to pay 2x more for r3.xlarge than for r3.large"
      • "Spot: each instance is worth number of vCPUs. Now, fetch me 1000 vCPUS"
• Can use different instance types (better chances of getting bid price on at least one type)
• Allocation strategy
  • lowestPrice - use spot pool with lowest price 
    • good for small fleets that are short running (minimizes cost)
    • "puts all your eggs in one basket"
  • diversified - distirbute across many spot pools 
    • good for large fleets that are long running (minimizes interruptions)
• Best practices (stock market)
  • Older generations have more stable pricing
  • Use multiple capacity pools
• Data feed
  • log of spot usage provided

Spot request (obsolete)

• spot price - current market price (lowest fulfilled bid)
• constraints (optional)
  • launch group
    • instances launched and terminated together 
    • if AWS has to terminate it will terminate whole group 
    • if you terminate yourself the rest stays
  • availability zone group     
    
    • all instances in the same group
    • if AWS has to terminate it will not terminate the whole group (unlike launch group)
• AZ selection
  • EC2-Classic: any AZ with lowest bid
  • Default VPC - chosen subnet and its AZ (or if not specified chooses subnet itself, but may not be the cheapest)
  • Nondefault VPC - chosen subnet
• Spot instance request
  • bid price, instance count, launch specification, one time | persistent, ValidFrom, ValidUntil
  • status: open, active (fulfilled), failed (bad parameters), closed (interrupted OR terminated), canceled

Spot bid (obsolete)

• how much (at most) you are willing to pay for an instance
• you get instance when your bid price > market price
• one-time
• persistent (automatically resubmitted when spot instance terminates)
• obsolete (no need to specify it anymore)

References

• https://www.youtube.com/watch?v=BlNPZQh2wXs
• https://aws.amazon.com/blogs/aws/new-resource-oriented-bidding-for-ec2-spot-instances/
• https://aws.amazon.com/blogs/aws/focusing-on-spot-instances-lets-talk-about-best-practices/
• https://aws.amazon.com/ec2/spot/bid-advisor/
• https://www.quora.com/When-you-use-an-AWS-EC2-instance-for-less-than-an-hour-are-you-charged-for-a-full-hour