47 - Observation is a read-only system - reference notes

Exercises 1-2: a metrics system, and proving it read-only

The metrics system is one vectorised pass: population = world.n_active, mean_e = world.energy[:n].mean(), append the row to a metrics array indexed by tick. Read-set: the world columns. Write-set: the metrics array, which nothing else touches.

The proof is a hash. Run 1,000 ticks with the metrics system in the DAG, hash the world; run again with it removed, hash again. They must be identical, because a disjoint write-set (§31) cannot perturb the world. If they differ, the observer wrote a column it should not have - find it. The simulator’s own inspect is exactly this system, and the determinism gate already proves it does not perturb (the run is bit-identical with it running).

Exercise 3: measure the thermometer

Time the tick with metrics on and off; vectorised reductions cost a fraction of a percent and the two are indistinguishable. Now make the metrics system np.sort(world.energy[:n]) each tick to report a median, and the reported tick time climbs - the thermometer is now heating the water. np.partition (O(N), no full sort) or a streaming quantile estimate recovers the budget. The Python-specific trap underneath: a metrics reduction written as a Python for loop is interpreter-bound and shows against the tick immediately; keep it in numpy.

Exercises 4-5: trace, and structured query

A trace is a filter, not new storage: query the §37 log for one entity id across the tick range and read its life off the events. A structured-log query - “every tick where population fell more than 10%” - is a vectorised diff over the population column; the same question over print() text is grep and hope, because the events were flattened to strings that no longer carry their types. A string is for one human reading one line; a structured event is for a system reading a million.

Exercises 6-7: an alert is a system; behind the queue

The alert reads the metrics table and emits a threshold crossing (population zero, tick-over-budget for T ticks) - one more read-only system, its read-set another system’s write-set. Shipping metrics out through the §35 queue and pausing the sink shows the inversion: the tick rate is unaffected and samples are dropped, not stalled. Backpressure on observability degrades observability, never the system.

Exercise 8: the guaranteed metric

A counter you must not lose - a billing total - is not fire-and-forget; it obeys the §46 rule. Do not advance the “reported” watermark until the sink confirms the batch durable, and on restart resend from the watermark. Contrast the cost: the fire-and-forget metric pays nothing and may gap a chart; the guaranteed one pays an fsync-class confirmation per batch and may never lose a count. The decision is per metric: a dropped frame-rate sample is a gap, a dropped charge is theft.