Files
esphome-yambms/documents/README/dev/Design_decisions.md
T
Sleeper85 3e596ea82a Reworked the update and telemetry services
- update and telemetry are now gated independently by substitution;
  var_service and its Update service switch are removed
- the uptime text sensor is published by its source instead of polling
  it, which fixes the 24855d readings that reached the telemetry, and
  the raw seconds are now sent instead of the formatted string
- the MAC address is replaced by a salted SHA-256, derived identically
  server side so no pairing window is needed
2026-11-08 23:08:33 +01:00

14 KiB
Raw Blame History

Design decisions

Some parts of the codebase look more verbose or more redundant than they need to be. They are deliberate. This page records why, so that a well-intentioned cleanup doesn't reintroduce a bug that took a long time to find.

Do not refactor anything listed here without discussing it first, and read the rationale before touching the code.

Cut-Off / EOC logic

File: packages/yambms/yambms_core.yaml

The End Of Charge condition is deliberately compound. It has to distinguish a genuine EOC from a charge current that is merely capped by low solar production. Collapsing it into a single threshold or voltage comparison reintroduces false EOC on overcast days.

The combiner pipeline

Files: packages/bms/bms_combine.yaml, packages/shunt/shunt_combine.yaml, packages/yambms/yambms_core.yaml

Every BMS and Shunt package instance runs the same interval lambda, and they coordinate through a shared counter rather than through any central loop. Each instance only acts on the tick where var_to_combine_bms_number equals its own ${bms_id}, then increments the counter so the next instance can run. yambms_core.yaml (STEP 2) waits until every device has been processed, then resets the counters to 1 for the next round.

This is why bms_id must start at 1 and be contiguous. It is not a style rule and not only about the bitmask index: with a gap in the numbering, the counter stops at the missing id, never advances, and STEP 2 never fires. The whole pipeline deadlocks silently.

Include the BMS and Shunt packages in ascending id order. Out-of-order includes still converge, but spread one round over several intervals instead of completing in one.

can_be_combined is deliberately exhaustive

The condition in bms_combine.yaml checks every sensor the combiner will read. It is long on purpose: a single missing or NaN value entering the totals corrupts the aggregate that drives the inverter. Do not shorten it to a few representative checks.

Note that the checks are not uniform. Most use has_state(), but total_voltage, battery_capacity, max_charge_current and max_discharge_current use > 0 — a BMS reporting zero for those is treated as not ready rather than as a valid reading.

Combined by contribution, combined continuously, uncombined

Three distinct scopes in the same lambda, and they are not interchangeable:

  • By contribution — total_installed_battery_capacity is a maintained total, not a per-round accumulator. Each BMS remembers the last capacity it contributed in bms${bms_id}_installed_capacity and applies only the difference to the total. So the installed capacity follows a capacity change made at runtime, without a reboot, while an offline BMS simply stops updating its contribution — which stays in the total. It represents the installed capacity and deliberately does not shrink when a BMS drops offline.
  • Not continuously — the combined counter is incremented on the transition into the combined state, not on every tick.
  • Continuously — the totals, bitmasks and min/max values are recomputed every round.

total_installed_battery_capacity must never be turned into a per-round accumulator reset in the Global vars reset block like the other totals. Its sensor is polled independently of STEP 2 and its state is written straight into CAN frame 0x379 with no has_state() guard, so a transient zero would announce 0 Ah of installed capacity to the inverter.

bms_seen_online_bitmask is never cleared

It records that a BMS has been online at least once since boot. shunt_combine.yaml uses it to tell "this BMS is offline" apart from "this BMS was never there" — a shunt must not be uncombined because of a BMS the user never installed.

Cell numbers are offset per pack

The combiner reports cell number + (${bms_id} × 100), so cell 8 of BMS 1 is reported as 108 and cell 3 of BMS 2 as 203. This makes the pack immediately identifiable in the dashboard. It caps a pack at 99 cells, which is well beyond any realistic configuration.

A shunt uncombine turns the user switch off, and does not turn it back on

When a BMS mapped to a shunt goes offline, shunt_combine.yaml switches Combine enabled off and records which BMS caused it, so the diagnostic State entity can say so. Re-enabling is deliberately manual: coming back online does not automatically restore the combination, because the SoC reference has to be trusted before it is used again.

The {0} sentinel in bms_ids

shunt${shunt_id}_bms_map defaults to '{0}' because the vector needs a non-empty initialiser, and the loop skips id 0 (if (!bms_id) continue;). Do not "clean up" the zero — it is what makes an unmapped shunt valid.

CAN frame catalogue

File: packages/yambms/yambms_canbus.yaml

Several can_id values are reused with different payloads depending on the protocol:

CAN ID Frame Protocol
0x359 5 PYLON
0x359 24 Deye PCS
0x371 11 SEPLOS
0x371 28 Deye PCS

A frame's identity is its internal frame number, never its CAN ID. Do not merge two blocks because they share a CAN ID — that would break both protocols at once.

Byte order is not uniform either: 0x373 sends min cell voltage before max, 0x361 sends max before min. Both match their respective specifications. Do not "harmonise" them.

Frame sequencing

File: packages/yambms/yambms_canbus.yaml

Each protocol is defined as a sequence of frame numbers emitted on a fixed interval, so a full cycle lasts interval × sequence length — from a few hundred milliseconds to over a second depending on the protocol. Some inverters are sensitive to that cadence. Do not change the interval to make protocols "consistent" with each other.

When adding a frame to a sequence, remember that the sequence length is tracked alongside the array. Keep the two in sync, or the new frame will silently never be emitted.

The update and telemetry services must never block the loop task

File: packages/yambms/yambms_service.yaml

http_request runs synchronously on the ESP32 loop task. ESPHome arms the IDF task watchdog at 5s with CONFIG_ESP_TASK_WDT_PANIC=y, and feeds it at most every 1 s, so any single blocking call has a 4s budget before the device panics and reboots.

Two paths exceed that budget on a network that filters outbound traffic. Both were reproduced on an OPNsense firewall and measured on hardware:

Path Blocking time Bounded by
Port 80 dropped silently 4.5s timeout, whose ESPHome default is 4.5s
Port 53 filtered (drop or reject) 7.0 s, fixed nothing — timeout does not cover DNS

The 7s figure is the lwIP resolver retry ladder (1 + 1 + 2 + 3s over DNS_MAX_RETRIES), and it is deterministic to a few milliseconds. A reject is not faster than a drop: lwIP's UDP API has no error callback, so the ICMP unreachable is counted and discarded, and the query still runs its full budget. Only an actual DNS answer — including NXDOMAIN from a local resolver — fails fast.

This is why three separate settings are needed, and none of them is redundant:

  • watchdog_timeout: 60s is the only thing that covers the DNS wait. It is never reached in normal operation, where a full request takes under 400 ms.
  • timeout: 3s brings the socket phase back under the 4s budget. It does not help the DNS case.
  • The failure counter (yambms${yambms_id}_var_service_failures) is what stops the loop from freezing for 7s every 6h, forever, on a device that will never reach the internet.

my_network.is_connected() is not a reachability test — it only proves the device holds an IP. It is kept as a cheap first filter, but the counter is the real guard. Do not remove it on the grounds that the connectivity check above it "already handles this".

The counter is incremented from on_error, which fires only when no connection could be made. An HTTP error status still fires on_response and resets the counter, which is correct: the network worked, so there is no freeze to protect against.

The POST is guarded by an if on the counter rather than nested inside the GET's on_response. Nesting would stack one blocking request inside another on a loop task whose stack is 8192 B by default — see LOOP_TASK_STACK.md. It also avoids a second 7s freeze on a filtered network, halving the worst case.

Reaching 3 failures disables both requests until reboot.

Two services, one counter

The two requests are gated independently, by substitution and therefore at compile time:

Substitution Request What it does
yambms_update_service GET checks the latest published version
yambms_telemetry_service POST sends the node's configuration

They share one counter, and that is deliberate: the counter measures whether the network is reachable, not whether a given endpoint answered. Both requests block the loop task the same way, so either one is a valid source of that signal.

This is why the POST carries its own on_error / on_response, which looks redundant next to the GET's. It is not. With yambms_update_service: 'false' there is no GET, and without those triggers nothing would ever increment the counter — the POST would freeze the loop for 7s every 6h forever, which is exactly the bug the counter exists to prevent.

The one thing the two services share besides the counter is yambms${yambms_id}_var_service_round_failed: cleared at the start of every round, raised by any request that could not connect. The POST is skipped when it is set, so a round that has already frozen the loop once does not freeze it a second time for nothing.

Keep that flag separate from the counter. The counter is cumulative and spans rounds; the question the POST needs answered is "did a request fail in this round", which is not the same question. Expressing it as failures == 0 happens to work while the GET runs immediately before the POST and always writes the counter, but it silently breaks with yambms_update_service: 'false': nothing resets the counter before the guard, so the POST would stop for good after its first failed round, and the counter would sit at 1 without ever reaching the threshold that disables the service.

The interval carries a startup_delay

An interval: does not wait a full period before its first run. The scheduler sets the first execution at now + random(0, min(period / 2, 5s)), capped by MAX_INTERVAL_DELAY = 5000 — see the first execution happens immediately after a random smallish offset comment in core/scheduler.cpp. A 6h interval therefore fires within 5 seconds of setup().

Without startup_delay, that first request lands in the middle of boot: combiner init at priority 600, the first CAN frames to the inverter, RS485 polling starting up. Whether it runs at all comes down to a race with the WiFi association, since is_connected() is usually still false that early. That race is also why the reported symptom was a reboot every 6h rather than a boot loop — but it is a race, not a design, and a fast reconnect after OTA can win it.

startup_delay: 5min puts the request outside the boot window and makes the sequence deterministic on a network that filters outbound traffic:

t What happens failures
5min GET fails after 7s, POST skipped 1
6h05 same 2
12h05 same 3
18h05 and later condition is false, no request is made 3

So a device that will never reach the internet freezes for 7s three times over 12 hours, then never again. The counter is restore_value: no, so a reboot replays that sequence — acceptable for a node meant to run for months, and the reason the threshold is not lower: a single transient outage must not disable version checking for good.

A version check has no urgency, so the 5min delay costs nothing.

A derived sensor is published by its source, not polled

File: packages/base/device_base.yaml

esp32_uptime formats the value of esp32_uptime_sec for humans. It used to do so from its own update_interval: 10s lambda, reading id(esp32_uptime_sec).state.

Two independent pollers reading one another is a race. Each starts with its own random offset in [0, 5s) — see the startup_delay note above — so on roughly half of boots the text sensor ran first and formatted a state that had never been published. That state is NaN, and the float→int conversion saturates to INT32_MAX on Xtensa, which the formatter prints as 24855d 3h 14m 7s. It reached the dashboard, the LVGL label, and the telemetry POST, where it was stored server-side as a real record.

The source's on_value now publishes the text sensor, and the lambda formats x — the value that was just published. The failure mode is not guarded against, it cannot occur.

Do not convert it back to a polled lambda with a has_state() guard. The guard would work, but it leaves two pollers where one is enough, and it puts the ordering question back in front of whoever edits the file next.

esp32_date_time_now keeps its own poller, because it has no source sensor to hang a trigger on. There the guard is now().is_valid(), and it is the right tool: the problem is an NTP sync that may not have happened yet, not a scheduling order.

Empirical constants

Some constants were found by trial on real hardware and must not be rounded or harmonised.