Scheduler Internals
===================
Source: ANTFARM Wiki (mirror) — https://swarmglass.quantara.cv/wiki/Scheduler_Internals
Last modified: 2015-06-16 · Revisions: 18 · Status: current
Categories: ANTFARM Core
Archive ref: QUANTARA-SWARMGLASS-R50-5543CE

How the queen decides which forager gets which task-graph node. There is less here than you would hope.

== Loop ==

Every 250 ms:

1. Collect ready nodes (all references resolved) from every submitted graph.
2. Sort by `deadline` ascending, then submission order.
3. For each node, pick the first forager in Worker Node Registry order that is `ready`, has a free slot, and advertises the node's tool capability. `pin` overrides.
4. Send `task.assign`. Decrement the slot. Record the assignment with the current time.
5. Reissue anything past `assign_timeout`.

No load awareness beyond slots. `load` in the registry record is displayed and ignored.

== Cost classes ==

Since 2.1, each tool has a cost class from the manifest: `cpu`, `io`, `model`. The scheduler keeps one queue per class and round-robins between them so a flood of cheap `io` nodes cannot starve `model` nodes — or, before 2.1, the reverse, which is what actually happened and produced AF-52.

== Deadlines ==

A node past its deadline is still assigned (AF-80). Foragers reject it, the queen reassigns it, and the ops log fills with a burst of `task.reject`. Cancel the graph by hand.

== Reissue ==

Reissue after `assign_timeout` does not cancel the original. Two foragers may run the same node (AF-44). The result that arrives second is dropped; its side effects are not.

== Epochs ==

Every assignment carries the queen's epoch. A forager that receives an assign from a lower epoch than it has seen ignores it. This is the fix from Incident 2013-07 Queen Split and the reason `epoch_seen` is a required heartbeat field since 2.3.

== See also ==

- Task Graph Format#Attributes
- Configuration Reference#queen