tiopatinhas (tp) is a companion for AWS's Auto Scaling. It attaches itself to an Availability Group and adds resources bought in Amazon's Spot market.
As we can lose those instances at any time (due to market conditions), tiopatinhas only allows itself to provide about 50% of the total number of running instances.
Other features include:
- efficient use of resources. tiopatinhas is aware of Amazon's one-hour billing cycles and has protections against "flapping"
- crash recovery: in case there is a market crash, tiopatinhas can acquire instances on the regular OnDemand market
- fail-safe design: tiopatinhas fails "up". If the process crashes, the worse thing that can happen is extra servers being left in the cluster. Amazon's Autoscaling rules are not modified at all and can take over at any time.
Our theory of operations is that even when the system is live, each server is actually about 50% idle so that the system can responde to changes in access patterns. How much you can commit to Spot Instances is correlated to how low you keep your load during normal system operations. Needless to say, you shouldn't use Spot Instances in systems which are not fault tolerant.
- Make sure the boto python package is installed in your system. If you use debian or ubuntu you can install it by typing: 'sudo pip install boto'.
- Make sure your AWS credentials are specified in a boto configuration file (typically ~/.boto). Instruction on how to setup this file can be found here: https://code.google.com/p/boto/wiki/BotoConfig
- Copy the template conf file (tp.conf.template) to tp/tp.conf so that the script can read it and make the changes according to your needs. Tio patinhas currently supports the following properties:
- max_price: A map that specifies the maximum bid prices for each type of EC2 instance. TP will use the prices specified in this map to bid for instances of that type in the spot market.
- max_candidates: The maximum number of instances TP will manage.
- instance_name: The prefix that will be used by TP to name managed instances.
- region: The AWS region where the AutoScaling group is located.
- placement: The AWS availability zone where TP instances will be launched.
- health_check_path: Server HTTP path that must respond to health checks before the TP instance is attached to the load balancer.
- lower_cpu: The AutoScaling CPU treshold for scaling down.
- upper_cpu: The AutoScaling CPU treshold for scaling up.
- spot_type The instance type to bid for in the spot market (recommended: the same instance from the ASG)
- emergency_type The instance type to buy in case of market crash (recommended: the same instance from the ASG)
- lower_treshold: The amount of measurements below the lower_cpu TP will consider before scaling down. (advanced)
- lower_treshold: The amount of measurements above the upper_cpu TP will consider before scaling up. (advanced)
- tags: A map containing custom metadata tags that must assigned to TP instances. (optional)
- user_data_file An optional script or data that must be supplied to the instance on startup. (optional)
- Once the tp/tp.conf file is ready, execute tiopatinhas issuing the following command:
- python tp.py -g <AutoScalingGroupName> (this command must currently be executed from within the "tp" folder)
- You must optionally supply options "-v" for verbose mode or "-d" for daemon mode.