bot-backend
4aec49f7fb
All checks were successful
Deploy Trade-In / changes (push) Successful in 12s
Deploy Trade-In / build-frontend (push) Has been skipped
Deploy Trade-In / build-browser (push) Has been skipped
Deploy Trade-In / test (push) Successful in 3m7s
Deploy Trade-In / build-backend (push) Successful in 59s
Deploy Trade-In / deploy (push) Successful in 1m13s
366 lines
16 KiB
Python
366 lines
16 KiB
Python
"""Scheduled backfill: detail-enrichment for legacy yandex listings (#1553).
|
||
|
||
Nightly window 12:00-15:00 UTC (migration 113, source=yandex_detail_backfill).
|
||
Offset from avito_detail_backfill (09-12 UTC) to avoid parallel egress on shared IP.
|
||
|
||
Problem: ~3952 yandex listings have detail_enriched_at IS NULL.
|
||
Yandex city sweep does not call YandexDetailScraper — SERP data only.
|
||
area_m2 coverage 23%, living/kitchen 0%, repair_state 1% on prod.
|
||
|
||
Solution: single snapshot SELECT at start (guarantees termination), fetch each
|
||
offer detail page via curl_cffi AsyncSession (chrome120 + proxy) — mirrors
|
||
yandex_address_backfill.py which already gets full HTML from Yandex on prod.
|
||
Parse HTML via YandexDetailScraper.parse (pure, no network). Persist via
|
||
save_detail_enrichment. Track consecutive parse→None results; abort after
|
||
max_consecutive_blocks. Прогон с нулём обогащений теперь 'failed', не 'done'
|
||
(#2674, runs.mark_backfill_finished): на проде 31 прогон из 52 упирался ровно в
|
||
этот брейкер (attempted=5 failed=5) и все 31 назывались успешными. Остаток
|
||
снапшота уедет в следующую ночь через NULL detail_enriched_at.
|
||
|
||
Почему брейкер срабатывал так часто (разобрано 2026-08-06, замеры в комментарии
|
||
у OFFER_URL_PATTERN): в очереди лежали карточки новостроек, у которых source_url
|
||
ведёт на сайт застройщика, а не на realty.yandex.ru/offer/<id>/. Парсер отвергает
|
||
такие URL регуляркой ДО сети — это не капча, а предрешённый parse→None. Идут они
|
||
пачками, поэтому «5 подряд» набиралось на первых же строках и обрывало прогон
|
||
целиком. Теперь снапшот-SELECT берёт только то, что парсер в принципе может
|
||
разобрать, а размер отброшенного видно в counters.unenrichable_pending.
|
||
|
||
Why curl_cffi and not YandexDetailScraper.fetch_detail:
|
||
fetch_detail uses BaseScraper._http_get (plain httpx, no proxy, no TLS
|
||
fingerprinting). On datacenter IPs Yandex returns captcha / shell-HTML
|
||
→ parse always returns None → backfill would be 0% effective. The
|
||
curl_cffi path (chrome120 impersonation + mobile proxy) is already proven
|
||
by yandex_address_backfill, which fetches identical offer detail pages.
|
||
"""
|
||
|
||
from __future__ import annotations
|
||
|
||
import asyncio
|
||
import logging
|
||
import time
|
||
from dataclasses import dataclass, field
|
||
|
||
from curl_cffi.requests import AsyncSession
|
||
from scraper_kit.providers.yandex.detail import YandexDetailScraper, save_detail_enrichment
|
||
from sqlalchemy import text
|
||
from sqlalchemy.orm import Session
|
||
|
||
from app.core.config import settings
|
||
from app.services import scrape_runs as runs_mod
|
||
|
||
logger = logging.getLogger(__name__)
|
||
|
||
__all__ = [
|
||
"OFFER_URL_PATTERN",
|
||
"YandexDetailBackfillResult",
|
||
"run_yandex_detail_backfill",
|
||
]
|
||
|
||
# Условие, при котором обогащение этого объявления вообще возможно (#2723-класс).
|
||
# `YandexDetailScraper.parse` первым делом ищет в URL `/offer/<цифры>/` и без него
|
||
# возвращает None ЕЩЁ ДО обращения к HTML (providers/yandex/detail.py:150) — то есть
|
||
# отказ предрешён регуляркой, а не капчей.
|
||
#
|
||
# Замер прода 2026-08-06: из 15 511 необогащённых yandex-объявлений 3 535 имеют
|
||
# source_url на сайт застройщика (macroserver.ru, prospect-federation.ru,
|
||
# strana.com, …) — так карточки новостроек ведут с выдачи Яндекса. Обогащено из
|
||
# них за всю историю 0; все 1 210 обогащённых — вида realty.yandex.ru/offer/<id>/.
|
||
#
|
||
# Вред не в бесполезности, а в том, что они идут ПАЧКАМИ (один свип — один
|
||
# застройщик) и упираются в брейкер «5 parse-None подряд», обрывающий ВЕСЬ прогон:
|
||
# 32 прогона из 53 закончились ровно так — attempted=5, enriched=0, 23 секунды.
|
||
# Плюс каждая такая попытка — запрос на чужой сайт, который мы всё равно выбросим.
|
||
OFFER_URL_PATTERN = "/offer/[0-9]+"
|
||
|
||
|
||
@dataclass
|
||
class YandexDetailBackfillResult:
|
||
"""Counters for one yandex detail backfill run."""
|
||
|
||
attempted: int = 0
|
||
enriched: int = 0
|
||
failed: int = 0
|
||
unenrichable_pending: int = 0
|
||
duration_sec: float = field(default=0.0)
|
||
|
||
def to_dict(self) -> dict[str, int]:
|
||
return {
|
||
"attempted": self.attempted,
|
||
"enriched": self.enriched,
|
||
"failed": self.failed,
|
||
"unenrichable_pending": self.unenrichable_pending,
|
||
"duration_sec": int(self.duration_sec),
|
||
}
|
||
|
||
|
||
async def run_yandex_detail_backfill(
|
||
db: Session,
|
||
*,
|
||
run_id: int,
|
||
params: dict,
|
||
) -> YandexDetailBackfillResult:
|
||
"""Backfill detail_enriched_at for legacy yandex listings via curl_cffi + parse.
|
||
|
||
Params (from default_params jsonb in scrape_schedules):
|
||
batch_size: int -- snapshot size (SELECT LIMIT), default 800.
|
||
budget_sec: float -- wall-clock budget per run, default 3600s.
|
||
request_delay_sec: float -- delay between listings, default 5s.
|
||
max_consecutive_blocks: int -- consecutive parse→None before abort, default 5.
|
||
|
||
Fetch mechanism:
|
||
curl_cffi AsyncSession(impersonate="chrome120") + scraper_proxy_url — mirrors
|
||
yandex_address_backfill. On HTTP 200: pass resp.text to
|
||
YandexDetailScraper().parse(html, offer_url). parse→None counts as a fail
|
||
(possible captcha wall); consecutive None → abort after max_consecutive_blocks.
|
||
|
||
Lifecycle: update_heartbeat -> snapshot -> loop with budget guard ->
|
||
mark_backfill_finished (done / failed при нуле обогащений, #2674);
|
||
mark_failed напрямую — только при исключении.
|
||
"""
|
||
batch_size = int(params.get("batch_size", 800))
|
||
budget_sec = float(params.get("budget_sec", 3600))
|
||
request_delay_sec = float(params.get("request_delay_sec", 5.0))
|
||
max_consecutive_blocks = int(params.get("max_consecutive_blocks", 5))
|
||
|
||
counters = YandexDetailBackfillResult()
|
||
current_counters: dict[str, int] = counters.to_dict()
|
||
|
||
start = time.monotonic()
|
||
|
||
try:
|
||
runs_mod.update_heartbeat(db, run_id, current_counters)
|
||
|
||
# SNAPSHOT: single SELECT at start -- NOT re-selected in loop.
|
||
# Priority: is_active DESC (active first), scraped_at DESC (newest first).
|
||
# Гейт по OFFER_URL_PATTERN — тот же признак, по которому парсер отказывает
|
||
# (см. комментарий у константы): в очередь не берём то, что заведомо
|
||
# непарсимо, иначе пачка карточек застройщика обрывает прогон брейкером.
|
||
snapshot = (
|
||
db.execute(
|
||
text(
|
||
"""
|
||
SELECT id, source_url
|
||
FROM listings
|
||
WHERE source = 'yandex'
|
||
AND detail_enriched_at IS NULL
|
||
AND source_url IS NOT NULL
|
||
AND source_url ~ CAST(:offer_url_pattern AS text)
|
||
ORDER BY is_active DESC NULLS LAST, scraped_at DESC NULLS LAST
|
||
LIMIT CAST(:batch_size AS int)
|
||
"""
|
||
),
|
||
{"batch_size": batch_size, "offer_url_pattern": OFFER_URL_PATTERN},
|
||
)
|
||
.mappings()
|
||
.all()
|
||
)
|
||
|
||
# Отброшенное не должно исчезнуть из виду: без этого счётчика «обогащено
|
||
# 12 тыс. из 15,5 тыс.» снова стало бы необъяснимым нулём (#2674).
|
||
counters.unenrichable_pending = int(
|
||
db.execute(
|
||
text(
|
||
"""
|
||
SELECT count(*)
|
||
FROM listings
|
||
WHERE source = 'yandex'
|
||
AND detail_enriched_at IS NULL
|
||
AND source_url IS NOT NULL
|
||
AND source_url !~ CAST(:offer_url_pattern AS text)
|
||
"""
|
||
),
|
||
{"offer_url_pattern": OFFER_URL_PATTERN},
|
||
).scalar_one()
|
||
)
|
||
if counters.unenrichable_pending:
|
||
logger.info(
|
||
"yandex_detail_backfill: run_id=%d — %d объявлений вне очереди: "
|
||
"source_url ведёт не на карточку Яндекса (%s), парсер их отвергает "
|
||
"до сети",
|
||
run_id,
|
||
counters.unenrichable_pending,
|
||
OFFER_URL_PATTERN,
|
||
)
|
||
|
||
if not snapshot:
|
||
logger.info(
|
||
"yandex_detail_backfill: run_id=%d -- no pending listings "
|
||
"(detail_enriched_at IS NULL = 0), done",
|
||
run_id,
|
||
)
|
||
# to_dict(), а не current_counters: пустая очередь при непустом
|
||
# unenrichable_pending — самый важный случай этого счётчика.
|
||
runs_mod.mark_done(db, run_id, counters.to_dict())
|
||
return counters
|
||
|
||
logger.info(
|
||
"yandex_detail_backfill: run_id=%d snapshot=%d (budget=%.0fs "
|
||
"delay=%.1fs max_consecutive_none=%d)",
|
||
run_id,
|
||
len(snapshot),
|
||
budget_sec,
|
||
request_delay_sec,
|
||
max_consecutive_blocks,
|
||
)
|
||
|
||
# Build proxies dict once — mirrors yandex_address_backfill.py
|
||
_proxy = settings.scraper_proxy_url
|
||
_proxies = {"http": _proxy, "https": _proxy} if _proxy else None
|
||
|
||
consecutive_none = 0
|
||
do_sleep = False
|
||
scraper = YandexDetailScraper()
|
||
|
||
async with AsyncSession(
|
||
impersonate="chrome120",
|
||
timeout=30.0,
|
||
proxies=_proxies,
|
||
headers={
|
||
"Accept-Language": "ru-RU,ru;q=0.9,en;q=0.8",
|
||
},
|
||
) as session:
|
||
for idx, row in enumerate(snapshot):
|
||
# Budget guard
|
||
elapsed = time.monotonic() - start
|
||
if elapsed > budget_sec:
|
||
logger.info(
|
||
"yandex_detail_backfill: run_id=%d -- budget %.0fs exhausted "
|
||
"(elapsed=%.1fs), stopping at #%d/%d",
|
||
run_id,
|
||
budget_sec,
|
||
elapsed,
|
||
idx,
|
||
len(snapshot),
|
||
)
|
||
break
|
||
|
||
# Delay before each request except the first
|
||
if do_sleep:
|
||
await asyncio.sleep(request_delay_sec)
|
||
do_sleep = True
|
||
|
||
source_url: str = row["source_url"]
|
||
listing_id: int = row["id"]
|
||
counters.attempted += 1
|
||
|
||
try:
|
||
try:
|
||
resp = await session.get(source_url, allow_redirects=True)
|
||
except Exception as fetch_exc:
|
||
consecutive_none += 1
|
||
counters.failed += 1
|
||
logger.warning(
|
||
"yandex_detail_backfill: run_id=%d listing_id=%d "
|
||
"fetch error (consecutive=%d): %s",
|
||
run_id,
|
||
listing_id,
|
||
consecutive_none,
|
||
fetch_exc,
|
||
)
|
||
if consecutive_none >= max_consecutive_blocks:
|
||
logger.error(
|
||
"yandex_detail_backfill: run_id=%d ABORT -- %d consecutive "
|
||
"errors. enriched=%d attempted=%d",
|
||
run_id,
|
||
consecutive_none,
|
||
counters.enriched,
|
||
counters.attempted,
|
||
)
|
||
break
|
||
continue
|
||
|
||
if resp.status_code != 200:
|
||
consecutive_none += 1
|
||
counters.failed += 1
|
||
logger.warning(
|
||
"yandex_detail_backfill: run_id=%d listing_id=%d "
|
||
"HTTP %d (consecutive=%d)",
|
||
run_id,
|
||
listing_id,
|
||
resp.status_code,
|
||
consecutive_none,
|
||
)
|
||
if consecutive_none >= max_consecutive_blocks:
|
||
logger.error(
|
||
"yandex_detail_backfill: run_id=%d ABORT -- %d consecutive "
|
||
"non-200 responses. enriched=%d attempted=%d",
|
||
run_id,
|
||
consecutive_none,
|
||
counters.enriched,
|
||
counters.attempted,
|
||
)
|
||
break
|
||
continue
|
||
|
||
enrichment = scraper.parse(resp.text, offer_url=source_url)
|
||
|
||
if enrichment is None:
|
||
# parse→None: captcha wall / shell-HTML / no JSON-LD.
|
||
# Do not mark listing as done — retry next night.
|
||
consecutive_none += 1
|
||
counters.failed += 1
|
||
logger.warning(
|
||
"yandex_detail_backfill: run_id=%d listing_id=%d source_url=%s "
|
||
"-> parse None (consecutive=%d)",
|
||
run_id,
|
||
listing_id,
|
||
source_url,
|
||
consecutive_none,
|
||
)
|
||
if consecutive_none >= max_consecutive_blocks:
|
||
logger.error(
|
||
"yandex_detail_backfill: run_id=%d ABORT -- %d consecutive "
|
||
"parse-None results (captcha wall?). enriched=%d attempted=%d",
|
||
run_id,
|
||
consecutive_none,
|
||
counters.enriched,
|
||
counters.attempted,
|
||
)
|
||
break
|
||
continue
|
||
|
||
consecutive_none = 0
|
||
if save_detail_enrichment(db, listing_id, enrichment):
|
||
counters.enriched += 1
|
||
|
||
except Exception as exc:
|
||
counters.failed += 1
|
||
logger.warning(
|
||
"yandex_detail_backfill: save/iteration error for listing_id=%d: %s",
|
||
listing_id,
|
||
exc,
|
||
)
|
||
try:
|
||
db.rollback()
|
||
except Exception:
|
||
pass
|
||
|
||
if counters.attempted % 25 == 0:
|
||
current_counters = counters.to_dict()
|
||
runs_mod.update_heartbeat(db, run_id, current_counters)
|
||
|
||
counters.duration_sec = time.monotonic() - start
|
||
current_counters = counters.to_dict()
|
||
runs_mod.mark_backfill_finished(
|
||
db, run_id, current_counters, source="yandex_detail_backfill"
|
||
)
|
||
logger.info(
|
||
"yandex_detail_backfill: run_id=%d FINISHED -- attempted=%d enriched=%d "
|
||
"failed=%d duration=%.1fs",
|
||
run_id,
|
||
counters.attempted,
|
||
counters.enriched,
|
||
counters.failed,
|
||
counters.duration_sec,
|
||
)
|
||
return counters
|
||
|
||
except Exception as exc:
|
||
counters.duration_sec = time.monotonic() - start
|
||
logger.exception(
|
||
"yandex_detail_backfill: run_id=%d FAILED after %.1fs",
|
||
run_id,
|
||
counters.duration_sec,
|
||
)
|
||
runs_mod.mark_failed(db, run_id, str(exc)[:1000], counters.to_dict())
|
||
raise
|