Main Loop Cadence Decoupled from Scheduler Wake Timing

May 14, 2026

J. Nick Koston

ESPHome Maintainer

Every component’s loop() now runs at the configured loop_interval_ (default ~62 Hz). Previously, when other scheduler activity was running — set_interval / set_timeout / PollingComponent updates with sub-loop_interval_ cadences — the component phase got pulled forward up to ~128 Hz; quiet configs with no such scheduler entries were unaffected and always ran at the documented rate. Background events (MQTT RX, USB RX, BLE events, ESPNOW, camera, micro_wake_word, speaker, USB host/CDC, lwIP socket) still wake their component within one tick via the existing wake_loop_* paths. App.set_loop_interval() — the documented knob for power savings — finally works.

This is a behavior change in ESPHome 2026.5.0 and later with no API break, but it affects every component whose loop() implicitly depended on running at ~2× the configured rate when scheduler activity was driving the pull-forward.

Background

PR #15792: decouple main loop cadence from scheduler wake timing

What was happening

Sampling runtime stats over 60 s on a typical ESP32 IDF config (api + esp32_ble + esp32_ble_tracker + bluetooth_proxy + debug + a couple of sensors) at the default loop_interval_ = 16 ms:

Component	Before	After	What’s counted
api	7650 / 60 s ≈ 128 Hz	3720 / 60 s ≈ 62 Hz	loop() calls
esp32_ble	7650 ≈ 128 Hz	3720 ≈ 62 Hz	loop() calls
esp32_ble_tracker	7650 ≈ 128 Hz	3720 ≈ 62 Hz	loop() calls
debug	7651 ≈ 128 Hz	3720 ≈ 62 Hz	loop() calls
bluetooth_proxy	589 ≈ 10 Hz	599 ≈ 10 Hz	100 ms set_interval (unaffected)

Every component with a real loop() was running at ~2× the configured loop_interval_. 62 Hz is what the documentation has always promised; the 128 Hz was emergent behavior.

Why it happened

Application::loop() bounded its sleep by min(loop_interval_ - elapsed, next_schedule_in()) with a delay_time / 2 floor. If the scheduler had any entry due sooner than loop_interval_ / 2, the loop slept for loop_interval_ / 2 and ran the entire component phase again — every component’s loop() got pulled forward by whatever else happened to be scheduling work. The effect was config-dependent and non-local: adding one scheduled item anywhere silently shifted every other component’s cadence.

Two recent trends made this bite harder: (1) more components are PollingComponents, each of which is a set_interval under the hood; (2) more components use set_interval / set_timeout directly for retries, debouncing, animations, and protocol timing. App.set_loop_interval() — the documented knob for power savings — was silently defeated by the same coupling.

Removing the floor is safe now because wake_loop_threadsafe() is accessible everywhere as of 2026.4.0 (see the 2026.4.0 wake_loop blog post). Any component that needs to wake the loop sooner than loop_interval_ has a proper way to do it without the floor papering over missing wake-ups.

What’s Changing

Application::loop() is restructured into two independent phases:

• Phase A (every tick): drain wake notifications, run scheduler.call(), feed the watchdog.
• Phase B (gated): iterate registered components. Runs when any of these is true:
  • loop_interval_ has elapsed since the last component phase, or
  • HighFrequencyLoopRequester is active, or
  • a background producer set the wake-request flag via wake_loop_* (MQTT RX, USB RX, BLE events, ESPNOW, camera, micro_wake_word, speaker, USB host/CDC, lwIP socket, or a component-level enable_loop_soon_any_context()).

Sleep between ticks is min(time-until-next-component-phase, next_schedule_in()). A scheduler-timer wake runs only Phase A; a wake_loop_threadsafe() wake runs Phase B too so the producer’s component can drain its queued work.

The delay_time / 2 floor is removed entirely. Zero-delay set_interval(0, ...) registrations no longer busy-loop the component phase, because Phase B is gated separately from scheduler timer expiry. HighFrequencyLoopRequester remains the correct mechanism for “I need fast wakes sometimes.” Zero-delay defer() is unaffected.

Ordering preserved: defer() → FIFO before components; scheduled items before components when both are due; scheduled callbacks still main-thread-only; the WDT is fed ≥ once per tick.

Who This Affects

Every component that overrides loop() — in-tree and external alike — but most won’t notice. Specifically:

• Components whose loop() depended on running at ~128 Hz — animations stepping per loop tick, debounce / state-machine timing keyed to loop count, anything that implicitly assumed the doubled rate. Those will now step at ~62 Hz when scheduler activity was previously providing the pull-forward.
• Background-event producers — already wired up. Every existing wake_loop_* caller (MQTT RX, USB RX, BLE events, ESPNOW, camera, micro_wake_word, speaker, USB host/CDC, lwIP socket, and per-component enable_loop_soon_any_context() calls) now sets a wake-request flag before signalling the platform so that Phase B runs in the next tick. No action needed in callers.
• Components that need sub-loop_interval_ cadence — see the migration guide below for the preferred event-driven pattern. HighFrequencyLoopRequester is available but should be a last resort.
• Power-managed configs using App.set_loop_interval() — this knob now actually saves power. If you previously bumped loop_interval_ to 100 ms and saw no power reduction, expect a real change this release.

Migration Guide

For most components, do nothing. Your loop() will simply run at the documented rate.

If your component genuinely needs to run more often than loop_interval_, the preferred pattern is event-driven, not high-frequency: keep the loop disabled while there’s nothing to do, and have your event source re-enable it. The Component base class has the primitives you need:

API	Where you call it	Purpose
this->disable_loop()	Main thread (typically the tail of loop() once your work is done)	Take the component out of the iteration list — its loop() won’t be called again until something re-enables it.
this->enable_loop()	Main thread	Put the component back in the iteration list.
this->enable_loop_soon_any_context()	ISR, FreeRTOS task, or any thread	ISR-safe enable. Sets a pending flag that the main loop drains on the next tick. There is intentionally no disable_loop_soon_any_context() — use disable_loop() from loop() itself.
App.wake_loop_threadsafe()	FreeRTOS task / non-ISR thread (BLE callbacks, network events, platform task callbacks)	Forces Phase B in the next tick globally; lets every enabled component drain queued work without changing per-component enable state.
App.wake_loop_any_context() / App.wake_loop_isrsafe()	ISR context	ISR-safe global wake. See Waking from ISR for per-platform IRAM and xHigherPriorityTaskWoken notes.

The typical event-driven pattern looks like this:

#include "esphome/core/application.h"
#include "esphome/core/component.h"

class MyEventDrivenComponent : public Component {
 public:
  void setup() override {
    // ... register hardware / callbacks ...
    this->disable_loop();  // nothing to do until an event arrives
  }

  void loop() override {
    this->drain_queued_events_();
    if (this->queue_empty_()) {
      this->disable_loop();  // re-enable from on_event()
    }
  }

  // Called from a FreeRTOS task / network callback / etc.
  void on_event(Event e) {
    this->queue_.push(e);
    this->enable_loop_soon_any_context();  // ISR-safe; also wakes the main loop
  }
};

enable_loop_soon_any_context() both flips the component’s enable bit and wakes the main loop, so loop() runs within a tick of the event landing — regardless of where loop_interval_ is set. This composes cleanly with App.set_loop_interval(...) for power-managed configs: idle current drops because the loop sleeps; latency stays low because events wake it.

WARNING

HighFrequencyLoopRequester is a last resort

HighFrequencyLoopRequester::start() keeps Phase B running every tick unconditionally. It exists for genuinely time-critical workloads (e.g. tight bit-banged protocol timing where you can’t wait for an event), and even there, scoping the request as narrowly as possible — start() only while the critical work is in flight, stop() immediately after — is the supported usage. If you’re reaching for it because “I just need faster wakes,” prefer disable_loop() / enable_loop_soon_any_context() first — it gets you the same latency at a fraction of the wakeup count.

The full architectural rationale, including how to decide between loop(), set_interval(), and the scheduler under the new cadence rules, is documented in Choosing Between loop() and the Scheduler.

Finding Code That Needs Updates

# Components with a loop() override (the work you might want to review).
# The narrower `override`-style pattern catches modern code; the wider one
# also picks up older external components that declared `void loop();` without
# `override` and is precisely where you're most likely to find code that
# implicitly depended on the old ~128 Hz cadence.
grep -rn 'void loop() override' your_component/
grep -rEn 'void[[:space:]]+loop[[:space:]]*\([[:space:]]*\)' your_component/

# Zero-delay set_interval calls — previously busy-looped the component phase
grep -rEn 'set_interval\(\s*0\s*,' your_component/

# Background producers that should be using a wake primitive
grep -rn 'wake_loop_threadsafe\|enable_loop_soon_any_context\|wake_loop_any_context\|wake_loop_isrsafe' your_component/

# Components already opted in to high-frequency wakes (no action needed)
grep -rn 'HighFrequencyLoopRequester' your_component/

If your loop() was implicitly relying on the ~128 Hz emergent rate and ~62 Hz is now insufficient, switch to HighFrequencyLoopRequester (for sustained fast wakes) or a wake primitive driven from your event source (for event-driven wakes).

Questions?

If you have questions about migrating your external component, please ask in:

• ESPHome Discord - #devs channel
• ESPHome GitHub Discussions

Related Documentation

• PR #15792 — decouple main loop cadence from scheduler wake timing
• PR #15846 — prerequisite WDT-feed rate-limit adjustment
• Architecture doc: loop control and main-loop cadence
• 2026.4.0 blog: wake_loop moved to core — provided the unconditional wake primitive that made this fix safe

FloatOutput Power Scaling Fields Gated Behind USE_OUTPUT_FLOAT_POWER_SCALING

Modbus Server Split Out of modbus_controller