How do I send video files that are being transcoded live with Python's Twisted.web?

Question

I'm building a video streamer using Python's Twisted.web with callbacks to ffmpeg.

I'm wondering, since I'm sending the files through using 'return file.read()' there'll be problems with sending the transcoded output since it won't be available to be read in entirety? Am I over thinking this and it's all managed by Twisted.web's low level management code or do I have to rethink my method?

Additionally, how can I support people skipping ahead to parts of the video that haven't buffered for them? I expect it to be a really complicated, but just incase it isn't.

Thanks

Answer 1

1) Can I use file.read() in twisted?

The short answer is "no".

Unless you've done a bunch of other relatively complicated things, such as set O_NONBLOCK on the fd, or explicitly delegated this request to separate thread, file.read() is a blocking call, that will read the entirety of the file into memory before it returns.

While this data is being read from disk synchronously, Twisted's reactor will be unable to invoke any callbacks or read or write any available data on any fds, causing any other pending requests or streaming transfers to grind to a complete halt while your file is being read into memory.

Moreover, if the source volume of the file is under heavy load and/or there are lots of files, even the open() calls to get file objects may block for a significant amount of time. There is no nonblocking open, so if you need a file object, you should use reactor.deferToThread() if you need to get a regular file object in twisted.

So here's what you should do:

For reading files directly from the filesystem with twisted, you should look into twisted.web.server.Site and twisted.web.static.File. There's an interesting tutorial on them here. Not only will these calls stream the file content from disk, but they will also support additional HTTP features for static files, such as HTTP Range requests.

2) How can I support people skipping ahead to parts of the video that haven't buffered for them?

This is commonly referred to as scrubbing, and there are basically two kinds of scrubbing: HTML5 scrubbing and time-based scrubbing.

The former (HTML5 scrubbing) is just a fancy way of saying "HTTP Range requests", and is supported by native browsers and plugins, such as Safari Mobile, Quicktime, mplayer, and all browsers that implement the html5 tag. This is supported by the twisted.web.static.File I mentioned in my answer to the first question.

The latter (time-based scrubbing) is the legacy way of supporting scrubbing. It is basically an agreement between a custom player (such as jwplayer or OSMF) and a server that allows the player to request the video starting at a time offset specified by a certain query parameter.

This can be done hackily by starting ffmpeg in a subprocess with the -itsoffset parameter pulled from the required query parameter, using some input file from disk, or using a module in your favorite webserver. Either way requires a player that understands what query parameter to send.