|
|
@ -1,6 +1,7 @@ |
|
|
|
from __future__ import unicode_literals |
|
|
|
|
|
|
|
import json |
|
|
|
import random |
|
|
|
|
|
|
|
from .common import InfoExtractor |
|
|
|
from ..compat import ( |
|
|
@ -156,9 +157,17 @@ class PluralsightIE(PluralsightBaseIE): |
|
|
|
format_id = '%s-%s' % (ext, quality) |
|
|
|
clip_url = self._download_webpage( |
|
|
|
request, display_id, 'Downloading %s URL' % format_id, fatal=False) |
|
|
|
# #6989: sleep 3 seconds to avoid 429 errors. |
|
|
|
# should help with #6842. |
|
|
|
self._sleep(3, display_id) |
|
|
|
|
|
|
|
# Pluralsight tracks multiple sequential calls to ViewClip API and start |
|
|
|
# to return 429 HTTP errors after some time (see |
|
|
|
# https://github.com/rg3/youtube-dl/pull/6989). Moreover it may even lead |
|
|
|
# to account ban (see https://github.com/rg3/youtube-dl/issues/6842). |
|
|
|
# To somewhat reduce the probability of these consequences |
|
|
|
# we will sleep random amount of time before each call to ViewClip. |
|
|
|
self._sleep( |
|
|
|
random.randint(2, 5), display_id, |
|
|
|
'%(video_id)s: Waiting for %(timeout)s seconds to avoid throttling') |
|
|
|
|
|
|
|
if not clip_url: |
|
|
|
continue |
|
|
|
f.update({ |
|
|
|