Why this page matters

This page explains how Timing Reports and Telemetry fits into the wider ZeroKernel execution model, what problem it is meant to solve, and what trade-off you are actually accepting when you use it in production firmware. The goal is not to treat Timing Reports and Telemetry as an isolated API call, but to understand where it sits inside bounded scheduling, queue discipline, fault visibility, and profile selection.

Read this topic as an operational contract. Start from the smallest working path, wire it into a lean profile first, and only expand into richer routing, diagnostics, or transport state after you can prove that the timing outcome is still worth the extra flash and RAM. That mindset is what keeps ZeroKernel useful on small boards instead of turning it into another bloated abstraction.

The safest pattern is always the same: define the runtime boundary, keep the hot path short, measure the effect with compare scripts, and only then scale complexity. The examples below are not filler; they show the smallest repeatable patterns you can lift into real firmware when you need clean integration instead of ad-hoc loops.

Three practical patterns

Core cadence pattern

Use one bounded task for the hot path, then let the scheduler keep the phase aligned over time.

    ZeroKernel.begin(boardMillis);
ZeroKernel.addTask("Fast", fastTask, 10, 0, true);
ZeroKernel.tick();

Deferred work pattern

Move non-critical routing and transport out of the immediate task body so fast paths stay predictable.

    const auto key = ZeroKernel.makeTopicKey("telemetry.sample");
ZeroKernel.publishDeferredFast(key, sampleValue);
ZeroKernel.flushEvents();

Runtime visibility pattern

Read the timing report and stats together so you can prove the cost of each abstraction layer.

    const auto stats = ZeroKernel.getStats();
const auto timing = ZeroKernel.getTimingReport();
Serial.println(timing.maxTickMs);

What to verify while you use it

Validate timing before you validate aesthetics. A cleaner API is not a win if fast misses rise.
Prefer the smallest profile that still matches the workload, then add optional modules only when the measured payoff is obvious.
Keep callbacks and transport steps bounded so watchdog, panic flow, and queue limits remain meaningful.

Common mistakes that make results misleading

Do not copy a demo pattern into production firmware without measuring it on the real board and real build profile you plan to ship.
Do not read success counters without reading queue depth, timing, and workload label next to them.
Do not enable heavier diagnostics and compatibility flags in a lean target just because the defaults looked convenient.

Recommended working sequence

Start from the smallest valid path

Boot the runtime, register the minimum useful task set, and prove that the baseline timing is clean before adding optional layers.

Add one layer, then measure it

Introduce routing, diagnostics, or transport one layer at a time so the cost and payoff remain obvious.

Publish only repeatable results

Update docs, charts, or public claims only after the same workload survives the same validation path more than once.

Core metrics

Metric	Why it matters
`fast_miss`	The clearest gate for deterministic workloads.
`fast_avg_lag`	Shows average schedule slip under a defined workload.
`fast_max_lag`	Shows the worst observed slip in a compare window.
`queue_max`	Shows whether deferred work is staying bounded.

Typical runtime pull

    const auto stats = ZeroKernel.getStats();
const auto timing = ZeroKernel.getTimingReport();

How to read transport metrics

Success and fail counts should be read together with total attempts and queue depth. A higher fail count can still be acceptable if throughput increased and the queue stayed bounded.

Timing and telemetry FAQ

Should I publish a benchmark from one run?

No. Use repeatable windows, consistent build profiles, and the same board target before publishing numbers.

Which number should I protect most aggressively?

For most embedded workloads, protect fast misses first. That is usually the most expensive failure to accept.

What is the safest way to validate this page on real hardware?

Start from the leanest profile that still matches the topic, run the narrowest compare script for this behavior, and only then move to heavier mixed workloads. Do not jump straight to a fully loaded build if the base timing is not yet proven.