Documentation
¶
Overview ¶
Package caps is the concurrency-caps model of design §8.1: which layers a job is counted against, what those layers hold now, and whether a job still fits. The model is pure — it reads no store; the engine builds a Usage from the live jobs and asks Room about each dispatch it considers.
Index ¶
Constants ¶
const ( LayerRunner = "runner" LayerAccount = "account" LayerModel = "model" LayerPlan = "plan" LayerGlobal = "global" )
The layers, in the order Room reports them full (§8.1's order).
Variables ¶
This section is empty.
Functions ¶
func HeadroomCap ¶
HeadroomCap is the account cap a task job sees on the foreman's account (§8.2): one lower than the configured cap, never below 0, so one slot stays reserved for the foreman. An uncapped account (nil limit) stays uncapped: there is no cap to lower.
func Room ¶
Room reports whether every capped layer has room for job (§8.1), and names the first one that does not: runner, account, model, plan, global. A foreman turn counts toward the runner, account and global caps only — the plan and model layers ignore it — so the foreman is never starved by its own tasks.
Types ¶
type Job ¶
type Job struct {
Runner int64
RunnerCap int // the runner's advertised max_parallel; 0 = unset
Account string // the agent's account label; "" = no agent recorded
Model string
Plan int64
Foreman bool
// Headroom marks a task job whose account is the foreman's (§8.2):
// the account layer sees HeadroomCap of the configured cap.
Headroom bool
}
Job is the dispatch Room is asked about: which runner would take the job, through which account and model, for which plan, and whether it is a foreman turn. Plan is 0 for a foreman turn.
type Limits ¶
type Limits struct {
Global *int // the engine's overall ceiling
Plan *int // max parallel tasks per plan
Runner *int // the config's runner cap; a runner may advertise its own
Account map[string]int // by account label
Model map[string]int // by model; a per-model cap such as 1 for a heavy model
Cost map[string]int // by "harness/model"
Tier map[string]string // by "harness/model"
// NoHeadroom turns the foreman headroom rule (§8.2) off. The rule
// is on by default, so the zero value keeps it.
NoHeadroom bool
}
Limits are the configured caps (§8.1): a job may start only if every layer on its path has room. A nil pointer, or a missing map entry, means the layer is not capped; a set limit of 0 caps the layer at nothing, which is different from not capping it.
Cost is not a limit but is configured beside them: the relative cost weight of one (harness, model) pair, which the foreman reads through `capacity` (§12.2). A missing entry there is the default weight of 1. Tier overrides the model-name table for every agent of a (harness, model) pair. Outcome record keys stay (repo, class, agent), so records earned under one label stay with the agent when an override changes its tier.
func (Limits) CostOf ¶
CostOf is the cost weight of one (harness, model) pair (§8.1): what the config named, else the default of 1.
type Usage ¶
type Usage struct {
Runner map[int64]int
Account map[string]int
Model map[string]int
Plan map[int64]int
Global int
}
Usage is the live-job count per layer (§8.1). Live means leased or running: a queued job holds no runner, no agent and no slot. An empty account or model — no agent recorded on the job — and a plan of 0 — a foreman turn — count at the layers that have them.
Source Files