EC2 Manager

Track, manage and alter the state of an EC2 account with a focus on spot requests. The application understands EC2 regions, instances and pools of instances. It does not understand Taskcluster Provisioning specifics like Capacity or Utility factors.

Architecture EC2-Manager is comprised of X major subsystems:

1. State Database
2. CloudWatch Event listener
3. Spot Instance Request Poller
4. Periodic Housekeeping
5. API

State Database

The state database is implemented in Postgres and has an interface provided by the ./lib/state.js file. This database is written to be as simple as possible while providing the data we need with transactional consistency where appropriate. Only the absolute minimal amount of information is stored in the database.

CloudWatch Event Listener

CloudWatch Events are the primary datasource for information on instance state. Information about instances is stored in the instances table. Whenever the state (e.g. pending, running) changes for an instance, a message is sent with the instance's id and the new state. When an event is received for the creation of an instance, we need to look up some metadata using the describeInstances EC2 API, but we unconditionally delete instance shutdowns. Whenever we get a message about an instance creation, we ensure that the spot request it was associated is removed from the list of spot requests we need to poll.

The messages from CloudWatch Events reach us through an SQS queue. In the case that the instance's metadata isn't available through the decribeInstances endpoint, we redeliver the message a number of times. If it is unsuccesful on the last attempt, we report it to Sentry and move on. The periodic housekeeping will ensure that it is inserted into the state database when appropriate.

Spot Instance Request Poller

Whenever a spot instance is requested from the API, we insert relevant metadata from it into the spotrequests table. Periodically the ec2-manager will check if any of the outstanding spot requests have been resolved. This is often not done because the request is fulfilled, thus generating a cloudwatch event, long before we poll for it. If this poller discovers a spot request which has entered a state where it will not be fulfilled, it is cancelled.

Periodic Housekeeping

Every hour, the EC2 Manager will request the EC2 API's view of the state. When doing this it will kill any instance which has exceeded the absolute maximum run time. For all other instances and spot requests, the state returned by the EC2 API will be used to confirm the view of state that the EC2 Manager has. Any instance in local state which is not in EC2 API state will be deleted from local state and any instance in EC2 API state but not local will be added to local state.

API

The API provided by EC2 Manager can be used manage EC2 instances. Of paricular note is that the endpoint for submitting spot requests requires a fully formed and valid LaunchSpecification.

Hacking

git clone https://github.com/taskcluster/ec2-manager
cd ec2-manager
yarn
yarn test

Deployment notes

When deploying, keep in mind the following:

1. SSH Pubkey used in the LaunchSpecification must match the one configured in the EC2 Manager. If you submit a LaunchSpecification with a different public key, it will be rejected.
2. This service is set up to auto deploy to the ec2-manager-staging Heroku app on pushes to the master branch. Deploying to production requires promotion