Playbooks are not a thing yet
They will provide a way to automatically schedule actions to perform orchestrated higher-lever operations across an entire cluster. Writing about playbooks as if they were already here makes the entire system (agents and core) easier to think about and design.
Aspects related to Playbooks and how they work are likely to change significantly.
Playbooks are data structures defined in code and cached in the primary store. When stored in the DB, playbooks should also have for some extra metadata such as a version or VCS commit ID or link to enable “playbook deploy” and origin tracking.
When execution of a playbook starts, a playbook run record is created in the primary store to track the overall state of the playbook and determine progression. The full playbook is also copied in the playbook run record so that concurrent changes to the playbook definition do not impact any already started run.
On top of state information for the playbook as a whole, playbook runs also have to store some per-node state, tracking details such as action result and “batch execution” style state. Agent state also identifies nodes that are added or removed from the cluster while the playbook was executing so playbooks can deal with these cases as appropriate.
This per-node state is stored in a dedicated collection on the primary store:
PENDINGto indicate it needs processing.
REMOVEDstates to indicate they need special consideration.
finish_tstimestamp is added to all records in the primary store (per-node as well as the “global” run state).