Root cause of the red PR #3494 CI job (9% progress, 75s life, no error line):
test_send_message_rate_limits_across_different_topics_same_chat mocked
asyncio.sleep as a pure no-op without advancing time.monotonic. The 3rd send
(over the test's limit=2) entered TelegramGroupRateLimiter.acquire(), which
recomputes wait_s from the real, unmocked clock every iteration - since the
fake sleep never advances it, the window never expires and the while-loop
busy-spins forever instead of actually waiting, until pytest-timeout kills it.
Fixed by advancing a fake monotonic clock inside fake_sleep, matching the
already-correct pattern used by the other tests in this file.
Also added _reset_telegram_shared_client (tests/conftest.py, same pattern as
_reset_estimate_rate_limiter): app.services.tgbot.shared._client is a
module-level singleton whose rate limiter otherwise accumulates real
wall-clock timestamps across the whole pytest session, not per test.
Documented honestly in config.py: the API-role budget is shared between
support web-chat mirrors and GlitchTip alerts with no priority between them,
so a large alert burst can make the web-chat wait out its own timeout and
return 502 - flagged as a known follow-up, not fixed here.
NOTE: a full `pytest -q --timeout=60` run still hangs further into the suite,
at tests/test_glitchtip_webhook.py::test_telegram_failure_returns_502_not_500.
Not root-caused within this session's budget - the test's _fake_telegram_client
fixture correctly monkeypatches glitchtip_module.get_telegram_client, but the
anyio worker thread running the ASGI request is seen parked in a real
event-loop poll/select wait, consistent with an actual (non-mocked) sleep
somewhere in that path. Needs a follow-up session with a fresh time budget.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JY6iWDnGDthdvsMWgK1BMG
Deep review of PR #3494 found the rate limiter unusable as designed:
- H1: acquire() waited unbounded even for interactive HTTP handlers (support.py,
glitchtip.py already pass a narrow `timeout` — reuse it as the queue wait cap
instead of editing those handlers, which are out of scope here). New
TelegramRateLimitedError (subclass of TelegramError) gives a fast, honest
502 instead of hanging past the caller's own budget.
- H2: the limiter is per-process (in-memory), but two processes write to the
same group (uvicorn API + bot worker) — giving each the same 18/min doubled
the platform ceiling. Split into telegram_group_rate_limit_api_per_minute
(12) and _bot_per_minute (6), sum kept below ~20.
- M1: bridge.py sends without an explicit timeout inherited "wait forever",
stalling the single-threaded poll loop (open DB session) past the SIGTERM
drain window. Bounded via rate_limit_max_wait=20s at the six call sites.
- M2: _locks/_sent_at grew unbounded on every unique DM chat_id. Added
opportunistic cleanup of fully-expired entries.
- L1: the "queue full" warning now logs once per acquire() call, not once
per sleep iteration.
- Corrected a factual error in the docstring: TELEGRAM_SUPPORT_CHAT_ID and
TELEGRAM_ALERTS_CHAT_ID are the SAME group on prod (topics differ only).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JY6iWDnGDthdvsMWgK1BMG
Бот падал молча на каждом сообщении, если тему форума удалили/переименовали:
узнавали об этом только по отсутствию сообщений у людей. tgbot_main теперь
один раз на старте проверяет getChat + typing-индикатор с message_thread_id
(единственный способ Bot API провалидировать message_thread_id без создания
видимого сообщения) и громко пишет error при отказе, не роняя процесс.
Второе: лимит Telegram (~20 msg/min) общий на всю группу, все темы делят
бюджет — всплеск GlitchTip-алертов вместе с потоком поддержки в ту же группу
уже давал 429 и терял сообщения. TelegramGroupRateLimiter — скользящее окно
per-chat_id (НЕ per-теме) с asyncio.Lock на чат, встроен прямо в
TelegramClient._request перед _post, поэтому считает все отправки независимо
от relay/прямого пути и без изменений в support.py/glitchtip.py (они уже идут
через общий клиент). Порог настраивается через
TELEGRAM_GROUP_RATE_LIMIT_PER_MINUTE (дефолт 18, чуть ниже потолка площадки).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JY6iWDnGDthdvsMWgK1BMG
