infra/stacks/f1-stream/files/backend/extractors/service.py
Viktor Barzin b1c21f78b9 f1-stream: drop broken curated, dedupe streams, accept all pitsport categories
User feedback: every stream on /watch shows ads but the player fails
to load. Three causes, three fixes:

1. CuratedExtractor's two hmembeds 24/7 channels (Sky F1, DAZN F1)
   sat at the top of the list and ALWAYS failed: they load the
   upstream's ad overlay then JW Player throws error 102630 (empty
   playlist; the obfuscated decoder produces no fileURL in our
   environment). Disabled the registration in extractors/__init__.py
   until/unless we find a working bypass — leaving the existing
   `CURATED_BYPASS = {"curated"}` shim in service.py so the swap is
   reversible.

2. Pitsport surfaces every WRC stage / MotoGP session as its own
   /watch UUID, but they all resolve to the same upstream m3u8 URL
   (e.g. RallyTV one master.m3u8 across all 22 Rally de Portugal
   stages). Added URL-keyed dedupe in service.run_extraction so the
   /streams response shows one row per actual stream.

3. The pitsport category filter was still narrowed to motorsport.
   Pitsport.xyz only lists curated sports broadcasts (WRC, MotoGP,
   IndyCar, NASCAR, Premier League Darts, Premier League football…),
   so the site's own selection is the right filter. Replaced the
   hand-maintained MOTORSPORT_KEYWORDS list with `bool(category or
   title)` — anything pitsport returns goes through. Streams that
   aren't actually live get filtered out downstream when the embed
   API returns an empty manifest.

Frontend: hls.js `lowLatencyMode` was on by default but RallyTV (and
most non-LL-HLS providers) don't ship the LL-HLS extensions, which
broke playback in real browsers. Default to `lowLatencyMode: false`.

Result: /streams is now 1 verified live entry (Rally TV WRC stage
currently airing); was 24 with the top 2 always broken + 22 dupes.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
2026-05-07 15:42:24 +00:00

270 lines
11 KiB
Python

"""Extraction service - manages extraction lifecycle: polling, caching, health checking, serving."""
import logging
from datetime import datetime, timezone
from backend.extractors.models import ExtractedStream
from backend.extractors.registry import ExtractorRegistry
from backend.health import StreamHealthChecker
from backend.playback_verifier import PlaybackVerifier
logger = logging.getLogger(__name__)
class ExtractionService:
"""Manages the extraction lifecycle: polling, caching, health checking, and serving.
Extraction runs on a background schedule (via APScheduler), never on
client request path. After extraction, health checks verify each stream
is live. Results are cached in memory, keyed by site_key.
GET /streams only returns streams that passed health checks, sorted by:
1. is_live (live streams first)
2. response_time_ms (fastest first)
"""
def __init__(self, registry: ExtractorRegistry) -> None:
self._registry = registry
# Cache: site_key -> list of ExtractedStream
self._cache: dict[str, list[ExtractedStream]] = {}
self._last_run: str | None = None
self._last_run_stream_count: int = 0
self._health_checker = StreamHealthChecker()
self._playback_verifier = PlaybackVerifier()
async def shutdown(self) -> None:
"""Release the headless browser instance owned by the verifier."""
await self._playback_verifier.shutdown()
async def run_extraction(self) -> None:
"""Run all extractors, health-check results, and cache them.
This is called by the background scheduler. Each extractor's
results replace its previous cache entry entirely. After extraction,
health checks are run to verify streams are live and measure
response times.
"""
logger.info("Starting extraction run...")
start = datetime.now(timezone.utc)
streams = await self._registry.extract_all()
# Dedupe by canonical URL — pitsport surfaces every WRC stage as a
# separate event but they all point at the same RallyTV master.m3u8
# (and similar for MotoGP weekend sessions). Keep the first
# occurrence so the user sees one entry per actual stream.
deduped: list[ExtractedStream] = []
seen_urls: set[str] = set()
for stream in streams:
key = (stream.embed_url or "").strip() or (stream.url or "").strip()
if not key or key in seen_urls:
continue
seen_urls.add(key)
deduped.append(stream)
if len(deduped) < len(streams):
logger.info(
"Deduped streams: %d -> %d (collapsed %d duplicate URL(s))",
len(streams), len(deduped), len(streams) - len(deduped),
)
streams = deduped
# Run health checks + headless-browser playback verification.
# Both stream types are now verified end-to-end so the user only
# ever sees streams that actually play in a browser.
if streams:
m3u8_streams = [s for s in streams if s.stream_type != "embed"]
embed_streams = [s for s in streams if s.stream_type == "embed"]
# m3u8 streams: cheap structural health check (validates manifest,
# checks first variant playlist), then a headless-browser test
# to confirm hls.js can decode and render frames.
if m3u8_streams:
stream_dicts = [s.to_dict() for s in m3u8_streams]
health_map = await self._health_checker.check_all(stream_dicts)
for stream in m3u8_streams:
health = health_map.get(stream.url)
if health:
stream.response_time_ms = health.response_time_ms
stream.checked_at = health.checked_at
if health.bitrate > 0:
stream.bitrate = health.bitrate
# tentatively mark live; final word comes from the verifier
stream.is_live = health.is_live
# Browser verification: applies to both m3u8 (only those that
# passed structural health) and embed (always — they have no
# other way to verify).
verify_items: list[tuple[str, str]] = []
for stream in m3u8_streams:
if stream.is_live:
verify_items.append((stream.url, "m3u8"))
for stream in embed_streams:
verify_items.append((stream.embed_url or stream.url, "embed"))
verdicts = await self._playback_verifier.verify_many(verify_items)
now_iso = datetime.now(timezone.utc).isoformat()
for stream in m3u8_streams:
if not stream.is_live:
continue # already failed health check
verdict = verdicts.get(stream.url)
if verdict is None:
continue # verifier disabled or unavailable
stream.is_live = verdict.is_playable
stream.checked_at = now_iso
# Curated streams skip the verifier — they are hand-picked
# 24/7 channels whose embed pages aggressively detect headless
# automation. We can't reliably confirm playback server-side,
# but we trust the curator. The user's real browser does NOT
# trigger the same anti-bot heuristics (real plugins, real
# mouse movements, etc.).
CURATED_BYPASS = {"curated"}
for stream in embed_streams:
stream.checked_at = now_iso
if stream.site_key in CURATED_BYPASS:
stream.is_live = True
stream.response_time_ms = 0
continue
key = stream.embed_url or stream.url
verdict = verdicts.get(key)
if verdict is None:
# Verifier unavailable — fall back to "trust extractor".
# This keeps the service usable even without playwright.
stream.is_live = True
stream.response_time_ms = 0
else:
stream.is_live = verdict.is_playable
stream.response_time_ms = verdict.elapsed_ms
# Group streams by site_key and update cache
new_cache: dict[str, list[ExtractedStream]] = {}
for stream in streams:
new_cache.setdefault(stream.site_key, []).append(stream)
# Replace cache for extractors that returned results.
# Clear cache for extractors that returned nothing (site went down, etc.)
for extractor_info in self._registry.list_extractors():
key = extractor_info["site_key"]
if key in new_cache:
self._cache[key] = new_cache[key]
else:
# Extractor returned nothing - clear its cache
self._cache.pop(key, None)
self._last_run = start.isoformat()
self._last_run_stream_count = len(streams)
live_count = sum(
1 for streams_list in self._cache.values()
for s in streams_list if s.is_live
)
elapsed = (datetime.now(timezone.utc) - start).total_seconds()
logger.info(
"Extraction run complete: %d stream(s) from %d extractor(s) in %.1fs (%d live)",
len(streams),
len(new_cache),
elapsed,
live_count,
)
def get_streams(self) -> list[dict]:
"""Return all cached streams as a sorted list of dicts.
Only returns streams that passed health checks (is_live=True).
Sorted by fallback priority:
1. is_live (live streams first) - filters to live only
2. response_time_ms (fastest first)
Returns:
List of serialized ExtractedStream dicts from all extractors,
filtered to live-only and sorted by response time.
"""
all_streams: list[ExtractedStream] = []
for streams in self._cache.values():
all_streams.extend(streams)
# Sort by fallback priority: live first, then fastest response
all_streams.sort(
key=lambda s: (not s.is_live, s.response_time_ms)
)
# Only return live streams to clients
live_streams = [s for s in all_streams if s.is_live]
return [s.to_dict() for s in live_streams]
def get_all_streams_unfiltered(self) -> list[dict]:
"""Return ALL cached streams including unhealthy ones.
Used for debugging and status endpoints. Sorted by fallback priority
but includes streams that failed health checks.
Returns:
List of all serialized ExtractedStream dicts.
"""
all_streams: list[ExtractedStream] = []
for streams in self._cache.values():
all_streams.extend(streams)
# Sort by fallback priority: live first, then fastest response
all_streams.sort(
key=lambda s: (not s.is_live, s.response_time_ms)
)
return [s.to_dict() for s in all_streams]
def get_streams_for_session(self, session_type: str) -> list[dict]:
"""Return cached streams filtered/annotated for a specific session type.
Currently returns all live streams (extractors don't yet differentiate by
session type). This method exists as a hook for future filtering,
e.g., some extractors might only have race streams but not FP streams.
Args:
session_type: The F1 session type (e.g., "race", "qualifying", "fp1").
Returns:
List of serialized ExtractedStream dicts (live only, sorted).
"""
# For now, all streams are potentially relevant to any session.
# Future extractors may tag streams with session types, at which
# point this method will filter accordingly.
streams = self.get_streams()
logger.debug(
"Returning %d stream(s) for session type '%s'",
len(streams),
session_type,
)
return streams
def get_status(self) -> dict:
"""Return extraction service status for the /extractors endpoint."""
extractor_list = self._registry.list_extractors()
extractor_statuses = []
for info in extractor_list:
key = info["site_key"]
cached = self._cache.get(key, [])
live_count = sum(1 for s in cached if s.is_live)
extractor_statuses.append(
{
"site_key": key,
"site_name": info["site_name"],
"cached_streams": len(cached),
"live_streams": live_count,
}
)
total_cached = sum(len(streams) for streams in self._cache.values())
total_live = sum(
1 for streams in self._cache.values()
for s in streams if s.is_live
)
return {
"extractors": extractor_statuses,
"total_cached_streams": total_cached,
"total_live_streams": total_live,
"last_run": self._last_run,
"last_run_stream_count": self._last_run_stream_count,
}