Merge pull request #103 from ekimekim/mike/cutter/full-cut

Resurrect non-experimental cut, now dubbed "full" (vs "fast") cut

common/common/segments.py

@@ -9,6 +9,7 @@ import itertools
 import json
 import logging
 import os
+import shutil
 import sys
 from collections import namedtuple
 from contextlib import closing
@@ -256,7 +257,9 @@ def streams_info(segment):
 
 
 def ffmpeg_cut_segment(segment, cut_start=None, cut_end=None):
-	"""Return a Popen object which is ffmpeg cutting the given segment"""
+	"""Return a Popen object which is ffmpeg cutting the given single segment.
+	This is used when doing a fast cut.
+	"""
 	args = [
 		'ffmpeg',
 		'-hide_banner', '-loglevel', 'fatal', # suppress noisy output
@@ -283,6 +286,22 @@ def ffmpeg_cut_segment(segment, cut_start=None, cut_end=None):
 	return subprocess.Popen(args, stdout=subprocess.PIPE)
 
 
+def ffmpeg_cut_stdin(cut_start, cut_end, encode_args):
+	"""Return a Popen object which is ffmpeg cutting from stdin.
+	This is used when doing a full cut."""
+	args = [
+		'ffmpeg',
+		'-hide_banner', '-loglevel', 'fatal', # suppress noisy output
+		'-i', '-'
+		'-ss', cut_start,
+		'-to', cut_end,
+	] + list(encode_args) + [
+		'-', # output to stdout
+	]
+	logging.info("Running full cut with args: {}".format(" ".join(args)))
+	return subprocess.Popen(args, stdin=subprocess.PIPE, stdout=subprocess.PIPE)
+
+
 def read_chunks(fileobj, chunk_size=16*1024):
 	"""Read fileobj until EOF, yielding chunk_size sized chunks of data."""
 	while True:
@@ -292,7 +311,7 @@ def read_chunks(fileobj, chunk_size=16*1024):
 		yield chunk
 
 
-def cut_segments(segments, start, end):
+def fast_cut_segments(segments, start, end):
 	"""Yields chunks of a MPEGTS video file covering the exact timestamp range.
 	segments should be a list of segments as returned by get_best_segments().
 	This method works by only cutting the first and last segments, and concatenating the rest.
@@ -362,3 +381,47 @@ def cut_segments(segments, start, end):
 			with open(segment.path) as f:
 				for chunk in read_chunks(f):
 					yield chunk
+
+
+def feed_input(segments, pipe):
+	"""Write each segment's data into the given pipe in order.
+	This is used to provide input to ffmpeg in a full cut."""
+	for segment in segments:
+		with open(segment.path) as f:
+			try:
+				shutil.copyfileobj(f, pipe)
+			except OSError as e:
+				# ignore EPIPE, as this just means the end cut meant we didn't need all input
+				if e.errno != errno.EPIPE:
+					raise
+	pipe.close()
+
+
+def full_cut_segments(segments, start, end, encode_args):
+	# how far into the first segment to begin
+	cut_start = max(0, (start - segments[0].start).total_seconds())
+	# how much of final segment should be cut off
+	cut_end = max(0, (segments[-1].end - end).total_seconds())
+
+	ffmpeg = None
+	input_feeder = None
+	try:
+		ffmpeg = ffmpeg_cut_stdin(cut_start, cut_end, encode_args)
+		input_feeder = gevent.spawn(feed_input, segments, ffmpeg.stdin)
+		# stream the output until it is closed
+		for chunk in read_chunks(ffmpeg.stdout):
+			yield chunk
+		# check if any errors occurred in input writing, or if ffmpeg exited non-success.
+		if ffmpeg.wait() != 0:
+			raise Exception("Error while streaming cut: ffmpeg exited {}".format(ffmpeg.returncode))
+		input_feeder.get() # re-raise any errors from feed_input()
+	finally:
+		# if something goes wrong, try to clean up ignoring errors
+		if input_feeder is not None:
+			input_feeder.kill()
+		if ffmpeg is not None and ffmpeg.poll() is None:
+			for action in (ffmpeg.kill, ffmpeg.stdin.close, ffmpeg.stdout.close):
+				try:
+					action()
+				except (OSError, IOError):
+					pass

cutter/cutter/main.py

@@ -16,7 +16,7 @@ from psycopg2 import sql
 
 import common
 from common.database import DBManager, query
-from common.segments import get_best_segments, cut_segments, ContainsHoles
+from common.segments import get_best_segments, fast_cut_segments, full_cut_segments, ContainsHoles
 
 from .upload_backends import Youtube, Local
 
@@ -267,7 +267,13 @@ class Cutter(object):
 
 		upload_backend = self.upload_locations[job.upload_location]
 		self.logger.info("Cutting and uploading job {} to {}".format(format_job(job), upload_backend))
-		cut = cut_segments(job.segments, job.video_start, job.video_end)
+
+		if upload_backend.encoding_settings is None:
+			self.logger.debug("No encoding settings, using fast cut")
+			cut = fast_cut_segments(job.segments, job.video_start, job.video_end)
+		else:
+			self.logger.debug("Using encoding settings for cut: {}".format(upload_backend.encoding_settings))
+			cut = full_cut_segments(job.segments, job.video_start, job.video_end, upload_backend.encoding_settings)
 
 		# This flag tracks whether we've told requests to finalize the upload,
 		# and serves to detect whether errors from the request call are recoverable.
@@ -550,6 +556,9 @@ def main(
 			This is useful when multiple upload locations actually refer to the
 			same place just with different settings, and you only want one of them
 			to actually do the check.
+		cut_type:
+			One of 'fast' or 'full'. Default 'fast'. This indicates whether to use
+			fast_cut_segments() or full_cut_segments() for this location.
 	along with any additional config options defined for that backend type.
 
 	creds_file should contain any required credentials for the upload backends, as JSON.
@@ -608,6 +617,7 @@ def main(
 	for location, backend_config in config.items():
 		backend_type = backend_config.pop('type')
 		no_transcode_check = backend_config.pop('no_transcode_check', False)
+		cut_type = backend_config.pop('cut_type', 'fast')
 		if type == 'youtube':
 			backend_type = Youtube
 		elif type == 'local':
@@ -615,6 +625,11 @@ def main(
 		else:
 			raise ValueError("Unknown upload backend type: {!r}".format(type))
 		backend = backend_type(credentials, **backend_config)
+		if cut_type == 'fast':
+			# mark for fast cut by clearing encoding settings
+			backend.encoding_settings = None
+		elif cut_type != 'full':
+			raise ValueError("Unknown cut type: {!r}".format(cut_type))
 		upload_locations[location] = backend
 		if backend.needs_transcode and not no_transcode_check:
 			needs_transcode_check.append(backend)

cutter/cutter/upload_backends.py

@@ -31,12 +31,14 @@ class UploadBackend(object):
 	The upload backend also determines the encoding settings for the cutting
 	process, this is given as a list of ffmpeg args
 	under the 'encoding_settings' attribute.
+	If this is None, instead uses the 'fast cut' strategy where nothing
+	is transcoded.
 	"""
 
 	needs_transcode = False
 
 	# reasonable default if settings don't otherwise matter
-	encoding_settings = [] # TODO
+	encoding_settings = ['-f', 'mp4']
 
 	def upload_video(self, title, description, tags, data):
 		raise NotImplementedError
@@ -59,7 +61,18 @@ class Youtube(UploadBackend):
 	"""
 
 	needs_transcode = True
-	encoding_settings = [] # TODO youtube's recommended settings
+	encoding_settings = [
+		# Youtube's recommended settings:
+		'-codec:v', 'libx264', # Make the video codec x264
+		'-crf', '21', # Set the video quality, this produces the bitrate range that YT likes
+		'-bf', '2', # Have 2 consecutive bframes, as requested
+		'-flags', '+cgop', # Use closed GOP, as requested
+		'-pix_fmt', 'yuv420p', # chroma subsampling 4:2:0, as requrested
+		'-codec:a', 'aac', '-strict', '-2', # audio codec aac, as requested
+		'-b:a', '384k' # audio bitrate at 348k for 2 channel, use 512k if 5.1 audio
+		'-r:a', '48000', # set audio sample rate at 48000Hz, as requested
+		'-movflags', 'faststart', # put MOOV atom at the front of the file, as requested
+	]
 
 	def __init__(self, credentials, hidden=False, category_id=23, language="en"):
 		self.logger = logging.getLogger(type(self).__name__)

restreamer/restreamer/main.py

@@ -13,7 +13,7 @@ import prometheus_client as prom
 from flask import Flask, url_for, request, abort, Response
 from gevent.pywsgi import WSGIServer
 
-from common import dateutil, get_best_segments, cut_segments, PromLogCountsHandler, install_stacksampler
+from common import dateutil, get_best_segments, fast_cut_segments, full_cut_segments, PromLogCountsHandler, install_stacksampler
 from common.flask_stats import request_stats, after_request
 
 import generate_hls
@@ -235,6 +235,9 @@ def cut(channel, quality):
 			if any holes are detected, rather than producing a video with missing parts.
 			Set to true by passing "true" (case insensitive).
 			Even if holes are allowed, a 406 may result if the resulting video would be empty.
+		type: One of "fast" or "full". Default to "fast".
+			A fast cut is much faster but minor artifacting may be present near the start and end.
+			A fast cut is encoded as MPEG-TS, a full as mp4.
 	"""
 	start = dateutil.parse_utc_only(request.args['start'])
 	end = dateutil.parse_utc_only(request.args['end'])
@@ -257,7 +260,14 @@ def cut(channel, quality):
 	if not any(segment is not None for segment in segments):
 		return "We have no content available within the requested time range.", 406
 
-	return Response(cut_segments(segments, start, end), mimetype='video/MP2T')
+	type = request.args.get('type', 'fast')
+	if type == 'fast':
+		return Response(fast_cut_segments(segments, start, end), mimetype='video/MP2T')
+	elif type == 'full':
+		# output as mp4 with no more specific encoding args
+		return Response(full_cut_segments(segments, start, end, ['-f', 'mp4']), mimetype='video/mp4')
+	else:
+		return "Unknown type {!r}. Must be 'fast' or 'full'.".format(type), 400
 
 
 def main(host='0.0.0.0', port=8000, base_dir='.', backdoor_port=0):