X-Git-Url: https://git.rapsys.eu/youtubedl/blobdiff_plain/415fdb62500dca2e22067a05008dfbf87c75b662..eeb287ed58128659e5de200bba1e0762fb972560:/youtube_dl/extractor/ooyala.py diff --git a/youtube_dl/extractor/ooyala.py b/youtube_dl/extractor/ooyala.py index 13f1282..d5b05c1 100644 --- a/youtube_dl/extractor/ooyala.py +++ b/youtube_dl/extractor/ooyala.py @@ -3,23 +3,36 @@ import re import json from .common import InfoExtractor -from ..utils import unescapeHTML +from ..utils import ( + unescapeHTML, + ExtractorError, +) class OoyalaIE(InfoExtractor): _VALID_URL = r'(?:ooyala:|https?://.+?\.ooyala\.com/.*?(?:embedCode|ec)=)(?P.+?)(&|$)' - _TEST = { - # From http://it.slashdot.org/story/13/04/25/178216/recovering-data-from-broken-hard-drives-and-ssds-video - 'url': 'http://player.ooyala.com/player.js?embedCode=pxczE2YjpfHfn1f3M-ykG_AmJRRn0PD8', - 'md5': '3f5cceb3a7bf461d6c29dc466cf8033c', - 'info_dict': { - 'id': 'pxczE2YjpfHfn1f3M-ykG_AmJRRn0PD8', - 'ext': 'mp4', - 'title': 'Explaining Data Recovery from Hard Drives and SSDs', - 'description': 'How badly damaged does a drive have to be to defeat Russell and his crew? Apparently, smashed to bits.', + _TESTS = [ + { + # From http://it.slashdot.org/story/13/04/25/178216/recovering-data-from-broken-hard-drives-and-ssds-video + 'url': 'http://player.ooyala.com/player.js?embedCode=pxczE2YjpfHfn1f3M-ykG_AmJRRn0PD8', + 'info_dict': { + 'id': 'pxczE2YjpfHfn1f3M-ykG_AmJRRn0PD8', + 'ext': 'mp4', + 'title': 'Explaining Data Recovery from Hard Drives and SSDs', + 'description': 'How badly damaged does a drive have to be to defeat Russell and his crew? Apparently, smashed to bits.', + }, + }, { + # Only available for ipad + 'url': 'http://player.ooyala.com/player.js?embedCode=x1b3lqZDq9y_7kMyC2Op5qo-p077tXD0', + 'info_dict': { + 'id': 'x1b3lqZDq9y_7kMyC2Op5qo-p077tXD0', + 'ext': 'mp4', + 'title': 'Simulation Overview - Levels of Simulation', + 'description': '', + }, }, - } + ] @staticmethod def _url_for_embed_code(embed_code): @@ -28,7 +41,7 @@ class OoyalaIE(InfoExtractor): @classmethod def _build_url_result(cls, embed_code): return cls.url_result(cls._url_for_embed_code(embed_code), - ie=cls.ie_key()) + ie=cls.ie_key()) def _extract_result(self, info, more_info): return { @@ -47,13 +60,30 @@ class OoyalaIE(InfoExtractor): player = self._download_webpage(player_url, embedCode) mobile_url = self._search_regex(r'mobile_player_url="(.+?)&device="', player, 'mobile player url') - mobile_player = self._download_webpage(mobile_url, embedCode) - videos_info = self._search_regex( - r'var streams=window.oo_testEnv\?\[\]:eval\("\((\[{.*?}\])\)"\);', - mobile_player, 'info').replace('\\"','"') - videos_more_info = self._search_regex(r'eval\("\(({.*?\\"promo\\".*?})\)"', mobile_player, 'more info').replace('\\"','"') + # Looks like some videos are only available for particular devices + # (e.g. http://player.ooyala.com/player.js?embedCode=x1b3lqZDq9y_7kMyC2Op5qo-p077tXD0 + # is only available for ipad) + # Working around with fetching URLs for all the devices found starting with 'unknown' + # until we succeed or eventually fail for each device. + devices = re.findall(r'device\s*=\s*"([^"]+)";', player) + devices.remove('unknown') + devices.insert(0, 'unknown') + for device in devices: + mobile_player = self._download_webpage( + '%s&device=%s' % (mobile_url, device), embedCode, + 'Downloading mobile player JS for %s device' % device) + videos_info = self._search_regex( + r'var streams=window.oo_testEnv\?\[\]:eval\("\((\[{.*?}\])\)"\);', + mobile_player, 'info', fatal=False, default=None) + if videos_info: + break + if not videos_info: + raise ExtractorError('Unable to extract info') + videos_info = videos_info.replace('\\"', '"') + videos_more_info = self._search_regex( + r'eval\("\(({.*?\\"promo\\".*?})\)"', mobile_player, 'more info').replace('\\"', '"') videos_info = json.loads(videos_info) - videos_more_info =json.loads(videos_more_info) + videos_more_info = json.loads(videos_more_info) if videos_more_info.get('lineup'): videos = [self._extract_result(info, more_info) for (info, more_info) in zip(videos_info, videos_more_info['lineup'])] @@ -65,4 +95,3 @@ class OoyalaIE(InfoExtractor): } else: return self._extract_result(videos_info[0], videos_more_info) -