Q&A — Phase 17.4 SchedulerCortex: EDF, Preemption, Prediction Integration

[web3guru888]

Q&A — SchedulerCortex (Phase 17.4)

Issue #443 · Show & Tell #444

---

Q1: Why EDF over a simple priority queue?

A pure priority queue (e.g. CRITICAL > HIGH > NORMAL) has no concept of when a task must complete. A NORMAL task with a 50 ms deadline will sit behind 100 HIGH tasks with 10-second deadlines and miss its window. EDF orders by deadline_ns first, dispatching the most time-constrained task regardless of nominal priority. Priority only breaks ties within the same deadline bucket — preventing indefinite starvation of CRITICAL tasks that share a deadline window.

---

Q2: How does preemption work in asyncio if we can't cancel a running coroutine?

AsyncSchedulerCortex does not own coroutine handles — it returns list[ScheduledTask] to CognitiveCycle, which actually executes the coroutines. Preemption works at the dispatch level: when a new task with an earlier deadline arrives, the currently-marked "running" task is flagged in _cancelled and re-enqueued. On the next run_tick(), the new task is dispatched first. The caller (CognitiveCycle) must implement cooperative cancellation: await asyncio.wait_for(task_coro(), timeout_ns/1e9). For true hard preemption, tasks should be wrapped in asyncio.Task and cancelled via task.cancel() with CancelledError handling.

---

Q3: What is the tick_ms tuning tradeoff?

tick_ms controls how frequently run_tick() fires. Lower values = tighter deadline tracking + more CPU overhead:

tick_ms	Scheduling Hz	Best for
10	100 Hz	Sub-100ms deadlines, latency-critical
50	20 Hz	Default — balanced cognitive cycle
200	5 Hz	Low-overhead background scheduler
1000	1 Hz	Maintenance-only tasks

For most ASI-Build cognitive modules (50–500 ms cycles), tick_ms=50 gives adequate resolution without spinning the event loop.

---

Q4: How does SchedulerCortex use PredictiveEngine outputs?

The prediction_loop() background task calls PredictiveEngine.predict(module) on each tick. When prediction.confidence > 0.8, it pre-submits a ScheduledTask with:

• earliest_start_ns = now + predicted_interval_ns — task cannot execute before the predicted activation time
• deadline_ns = earliest_start_ns + 5_000_000_000 — 5-second grace window
• prediction_basis = prediction.prediction_id — traceability link back to the specific prediction

This means CognitiveCycle dispatches the task right as it becomes due rather than queuing it reactively after the module fires.

---

Q5: deadline_miss_policy DROP vs DEMOTE — when to use each?

DROP (default): Stale tasks are evicted. Use when executing an expired task would cause inconsistency — e.g. a SafetyFilter.evaluate() based on observations that are now outdated. Better to miss and wait for the next cycle than act on stale data.

DEMOTE: Task is re-enqueued at TaskPriority.BACKGROUND with deadline += 60 s. Use for best-effort work that is still useful even if delayed — e.g. a KnowledgeGraph.consolidate() that doesn't have real-time requirements. DEMOTE prevents work loss while ensuring it doesn't crowd out urgent tasks.

---

Q6: How does SchedulerCortex integrate with GoalRegistry and PlanExecutor?

Integration flow:

GoalRegistry.get_active_goals()
  → GoalDecomposer.decompose(goal)
  → PlanExecutor.plan_steps()
  → for each step:
      cortex.submit(ScheduledTask{
          module  = step.module,
          action  = step.action,
          priority = TaskPriority.HIGH if goal.is_critical else NORMAL,
          deadline_ns = goal.deadline_ns,
          payload = step.parameters,
      })

SchedulerCortex thus acts as the execution buffer between high-level planning (GoalRegistry/PlanExecutor) and low-level dispatch (CognitiveCycle). Goal deadlines propagate directly to task deadlines.

---

Q7: Grafana — queue depth + deadline miss rate panels

# Queue depth + deadline miss rate dashboard
panels:
  - title: Scheduler Queue Depth
    type: timeseries
    targets:
      - expr: asi_scheduler_queue_depth
        legendFormat: Queue Depth
    fieldConfig:
      thresholds:
        steps:
          - value: 0;   color: green
          - value: 500; color: yellow
          - value: 800; color: red

  - title: Deadline Miss Rate (per min)
    type: timeseries
    targets:
      - expr: rate(asi_scheduler_deadline_misses_total[1m]) * 60
        legendFormat: "{{module}} ({{policy}})"
    fieldConfig:
      thresholds:
        steps:
          - value: 0;  color: green
          - value: 5;  color: yellow
          - value: 20; color: red

  - title: Preemptions / 5m
    type: stat
    targets:
      - expr: increase(asi_scheduler_preemptions_total[5m])
        legendFormat: Preemptions