The heart of the plugin is the agent scheduler. This is the central piece of plugin state, and it keeps track of cluster and agent state as the plugin runs.
The scheduler maintains one open client for each configured Aurora URL. Clients communicate with the cluster using Thrift over HTTP.
Only one thread at a time may use each client, and any error will cause the client to be closed as a precaution. Future calls will automatically open a new client to the cluster.
For each cluster, the plugin will check on the resource quota available to the configured role. When considering whether to launch new agents, it will try to avoid exceeding the quota. Operators should use the quota as a way of scaling the amount of resources available to the agents.
Each agent goes through a sequence of states during its lifespan. The scheduler's primary responsibility is to apply logic to the agents and manage their state transitions in response to events.
See the source file for commentary on individual states and transitions. This diagram can be regenerated from the source using graphviz:
$ dot -Tsvg < doc/agent-states.dot > doc/media/agent-states.svg