What's New in HTTP Live Streaming

Session 504 WWDC 2016

HTTP Live Streaming offers a reliable media playback experiences over a wide variety of network conditions. Join us to learn how to deliver offline playback using the same media assets you already host for online playback. See new additions to the HLS protocol, such as support for fragmented MP4 and in-playlist metadata.

[ Music ]

Is this thing on?

[ Applause ]

Good afternoon.

I'm Roger Pantos.

This is what's new in HTTP Live Streaming.

We have been talking about HLS now for 7 to 8 years.

And yet, every time I do one of these, marketing says.

"You've got to tell the people what's going on!"

And so here we go.

Here's what's new.

But, first, HLS in 20 seconds.

So HLS is about playing playlists.

What's a playlist?

It's one of these things.

It's a TXT file.

It's got tags, which are these blue things here.

And it's got segments, which are those white things.

Each one is 10 seconds of media.

If you want to play one, then you go rummaging through the tags looking for something interesting like I don't know this one says this is an INF segment.

So, load it first.

So you do that.

Then you load segment1.

You play it.

You load segment2.

You play it.

Maybe it's a live playlist.

So you refresh the playlist.

You load segment3.

You play that.

And that is basically HLS in a nutshell.

So now that we're all on the same page about that, let's go on to what's new.

Sweet. OK.

MPEG-4 Fragment Support.

Who is this MPEG-4 guy and why is he fragmented?

Actually, most of you are probably familiar with MPEG-4 files.

If you've got a movie on your hard disk, it's probably in MP4 format or one of its close cousins.

And if you've ever looked into it, it consists of one sample table usually it's at the beginning and then all the sample data following that.

And we did that originally.

Then, we introduced a new segment a little bit later called fragments.

And fragments are basically a way to take your MPEG-4 file and divide it up into these little islands of playable data.

A little sample table, a little sample data.

A little sample table, a little more sample data.

And these little islands are all called "Fragments".

So what we're announcing today is essentially that what you can do now is you'll be able to take one of these things, these MPEG-4 fragments, and use it as a segment in your HLS playlist just like you can do with transport stream files or elementary audio streams.

And we're updating the spec to sort of give you the details of that.

There should be a link to the session I think now.

But what it allows you to do is essentially use fMP4 segments, and you can do anything you can do with transport streams with fMP4.

That includes iframe playlists.

That includes discontinuities.

That includes live streams.

All that stuff.

Same set of features.

And of course, we're delivering it on all of our HLS platforms so you can choose to encode all your streams as fMP4 and deliver them everywhere we are.

Why would you do that?

Well, in a word, interoperability.

If you're in a situation where you have to encode your entire media library one way to deliver it to Apple's platforms, and then maybe you have to encode it again in another way to deliver it to Android for instance, that kind of sucks.

But if you could have a single library, then your delivery costs would go down, your storage costs would go down, you would get better cache utilization because you're no longer competing with yourself on the edge, and as sort of fringe benefit.

Also, if you've got a single production chain, you can use the same set of tools and validation across all of your ecosystems.

And as a fringe benefit, because fMP4 doesn't use the same kind of packetization that transport streams do, you get slightly more efficiency at lower bit rates.

So that's sort of a gimmie.

So how does that change HLS for you, the playlist author?

Well, as it turns out, not very much.

If we take our playlist here that we were looking at earlier and we convert it to fMP4, all we have to do is this.

We change the initialization segment to point to a movie box instead, and then each segment becomes an MPEG-4 fragment.

And that's it.

Everything else is pretty much the same at the playlist level.

On the segment level, there's one other thing I was going to talk about, which is encryption.

Now, as you know, HLS has two different ways of encrypting your stuff.

You can either just encrypt the entire segment, AES CBC the entire thing.

And that's the same whether it's transport streams or fMP4.

For sample encryption, like if you want to use FairPlay Streaming, it's a little bit more complicated because you have to say what part of the files are encrypted and what part aren't.

And this time around last time, we did it for transport streams.

We essentially made up our own format.

This time, we decided to use an existing one.

And so this is what MPEG calls "Common Encryption".

It's a standard.

It came out a couple years ago.

The newest edition of the standard has a mode called cbcs, which is compatible with FairPlay Streaming.

And so if you're doing sample encryption in HLS, you'll be using common encryption cbcs mode for that.

And so with encryption, and playlist, and everything else, we pretty much have a whole story put together for HLS around fragmented MPEG-4.

But if we're going to get to the happy, shiny, bouncy world where everybody can just have a single copy of their catalog and deliver it everywhere, we have to go beyond HLS.

And one of the things we need to do is we need to write down a set of rules that says, "Here's how you're going to author your fMP4 segments or fragments, or what have you, so that they play everywhere."

And we've been working on that as well.

It started as sort of a what-if exercise between us and Microsoft.

And once we got to the point where it seemed like it was practical, it seemed like it would work, we took it and proposed it to MPEG.

And a lot of folks there seemed to think it was a really good idea.

So what it comes down to is it's a set of constraints for how you construct your fMP4 segments.

So you've got to deliver your audio and your video separately.

You've got to put an iframe at the start of every video segment.

You have to make sure you rigidly align your segments across your different bit rates.

You don't have to do any of this stuff if all you want to do is target HLS.

But if you want the broadest audience for that copy of your media library, then we're putting together a set of recommendations for you.

And so I think that that will end up benefiting everyone at least I hope so.

So now I'm going to switch gears a little bit and talk about something else everyone's favorite topic metadata.

So to sort of frame this, I'm going to start by talking about some of the existing things we already do for you in HLS for metadata.

Static metadata and ID3.

Static metadata is used for static metadata.

No surprise.

So it'll offer for like a content title, or something like that.

Most of it is text, and that makes it really easy to put into the playlist file either directly as a tag, or you can package it as a JSON and refer to it by URL.

Now you only get one title, or one author, or one copyright per presentation it's static.

But the benefit of that is that it's available whenever you want it.

It's available right away.

So let's contrast that to ID3.

Because timed metadata is linked to the timeline of the media presentation, it's often used for signaling events in the stream.

For instance, the start of an ad, the start of a program.

It lives inside the media format itself, so it's a binary format called ID3.

And that makes it compact.

It also makes it kind of difficult to author because it requires specialized tools.

It's kind of a pain in the butt if you're debugging it.

It's hard to read ID3.

But you can have as many as you want, and you can put them anywhere on the timeline you like.

And so that's powerful.

The flipside of that however is because it's in the media, we can't give it to you until it's actually been played.

So where does the new guy fit?

What we're adding is a way to author timed metadata inside the playlist itself.

And so you can use it for the same things that you're using ID3 for today.

You can author them as text.

So that's pretty easy.

You can easily put them into a playlist, and you can have as many as you want.

They can be overlapping.

They can nest.

They can do whatever.

And on the other hand, because they're in the playlist, we can give you the entire set as soon as we read the playlist no matter where the playhead is in the media.

And so that makes it pretty nice.

For doing things particularly like navigation control.

So what does it look like?

How do we make it sort of sit?

Well, we already had the idea of dates in HLS.

We have the PROGRAM-DATE-TIME tag that allows you to precisely signal times inside a playlist.

So we built on top of that and we said, each bit of metadata is going to be expressed as a range of time that has a set of attributes attached to it.

And some of those are ours.

Some of those are yours.

We set up the syntax so it'll be really clean to add and remove these things for live streams.

So it meshes really well with the live workflow.

So what does it look like?

Let's take a look at one of these guys.

We called the new tag the DATERANGE tag.

Here's another playlist.

The first guy says basically what time it is.

The second one is a DATERANGE tag.

And if you look at it, what you can see is it's composed of attributes.

These guys here, these blue ones, are the ones that are defined by the spec the ID, the START-DATE, the DURATION.

These guys, we reserve the X-namespace somewhere to HTTP for all you guys.

So you can use that to put in whatever attribute value pairs you want.

In this case, someone decided to put in an AD-ID and a beacon URL.

So that's an example.

That's what one of these guys looks like in a playlist.

When you're doing your content authoring, you can use it either live or VOD.

It just has to have a DATERANGE tag and a date.

You can obviously put these things in when you're creating your playlist.

But the other nice thing is, because the playlists are easy to manipulate, if you've got something like a post-production workflow that's doing ad insertion or something like that, it's really easy to augment the playlist and presentation with metadata at the same time.

By popular request, we also included rules for how to map SCTE-35 that you might find in your source media into the DATERANGE tags.

So you can carry SCTE-35 losslessly in HLS.

And finally, we've added support for the media stream validator.

And I actually wanted to take a moment to give a shout-out to the new set of tools we have.

We put together a talk we don't have time to talk about it right now in this session but we put together a talk talking about the new tools.

And so I'd encourage you to go watch that.

It's on demand on your apps or whatever you guys have.

And you should check it and out see what kinds of new tricks the validator in particular has learned.

So for playback, when you want to actually act on the metadata that's in the stream, you have a set of AVFoundation APIs.

And so you have to be in an app at this point.

And the APIs are essentially, give me all the data you currently got.

And then if you've got a live playlist, tell me when something new shows up.

And it's centered around a new object called the AVPlayerItemMetadataCollector.

And so we've got a little sample up here.

You can see it doing your typical things.

You create your asset.

You create your PlayerItem.

And then you say, "Hey, I want to collect some metadata."

So you create the MetadataCollector.

You set yourself up as a delegate so we can tell you stuff, and then you add it to the PlayerItem.

And that's it.

Then you get all the metadata items.

And so it's really easy to use.

It's really powerful.

And I think that it's going to be the preferred way to carry metadata in HLS pretty soon.

So the next thing we're going to talk about today is another feature that's come by popular request.

And that is the ability to play HLS when you're offline.

[ Applause ]

And to talk about that, I'm going to ask my colleague, Jordan Schneider, to come up and explain it to you.

[ Applause ]

Hey. So, let's see if those slides click.

There we go.

In iOS 10, we are bringing you the ability to download and then play HLS content without a network connection.

So now you can bring offline media playback to your users using your existing streaming video library.

So as part of this feature, we are extending the FairPlay Streaming support that we introduced to you last year to work with offline HLS, and we are providing you a way to download your HLS content even while your app is backgrounded.

And finally, we are exposing the ability to play partially-downloaded content even while your download might still be in progress.

So the first question - when and should you use offline HLS?

Well, yes, if you want to use this feature primarily when your user might want to play content when they expect that they might not have a network connection, such as when they go into airplane mode.

However, this feature is not targeted toward preemptively loading media.

We have other ways of doing that using AVPlayerItem.

So why use HLS for offline content rather than just downloading movie files?

Well, for one, your content might already be authored as HLS.

So adopting HLS for offline content might be a really convenient story for you from an engineering perspective.

But also, in addition to a video track, movie files can contain many different audio and subtitle tracks that your user might not really need or want to persist offline.

So in HLS, because all these tracks are contained separately on the server into different playlists, we can have a little bit more flexibility.

Specifically, we can have precise control over which media selections are downloaded, saving your user time, network data, and storage space on their device.

So, for example, by default we downloaded a user's default media selection, which is likely all they're really going to care about when they play offline.

What we're also providing you is a way to configure exactly what you and your user want to download to have available for offline playback.

So another cool thing about playing HLS offline is that, using the same asset, we can still download media selection options that have not been downloaded so long as they are still available on your server, which is nice.

So how do you do this?

In iOS 10, we're introducing a new class to download AVAssets called AVAssetDownloadTask.

This class inherits features of URL session.

Most importantly, the ability to download assets in the background even while your app isn't running using all of its mechanisms.

As I mentioned before, we're giving you the ability to control which media selections you want to download.

And because this is HLS, we have the ability for you to select the quality of your asset that you want to download for offline as well.

So the interface of AVAssetDownloadTask looks like this.

It inherits from URLSessionTask.

I want to mention here that it does not inherit from URLSessionDownloadTask.

It behaves similarly, but there is a few differences that I'll point out as the presentation goes along.

So to create one of these things, you want to set up a URLSession.

Now we have a specific subclass of URLSession called AVAssetDownloadURLSession that you have to use for one of these things.

And so you call makeAssetDownloadTask to create an asset download task.

And then for your selection of quality, we have a minimum required bit rate options key, as well as a media selection key.

Now I should note here that each AVAssetDownloadTask corresponds to a single set of media selections.

So if you want to download multiple media selections on the same asset, then you're going to want to set up multiple AVAssetDownloadTasks to do so.

I'll show you how to do that in a few slides.

So to use one of these things, here's what we do.

The first thing is to set up an AVAssetDownloadTask, configure it how you want it, and start the download.

Next, we want to respond to any events that might occur during the download, such as monitoring the progress of the download.

Then once the download finishes, we're going to want to store the location for playback when we're actually offline.

And then you might want to download additional media selections for your user to use.

And finally, you're going to want to play this thing.

So to set up one of these downloads, the first thing after you have your asset set up is you're going to want to set up a backgrounConfiguration from a background URLSessionConfiguration.

Then you're going to create your AssetDownloadURLSession, and then create your download task.

Here, I have the download task to download the quality at about a media bit rate of about two megabits per second.

But by default, we'll download the highest-quality video available.

Then once you have that set up, because this is a URLSessionTask, you just call resume to set your download.

So once you have your download going, then you're going to want to monitor it.

To do this, we have a new protocol for you called AVAssetDownloadDelegate.

So this inherits the same delegate methods that URLSessionTaskDelegate has.

But I want to point out two things that we have on here that we are introducing to you.

One is our method that we use to monitor the progress of downloads.

In this method, we express our progress of the download and time ranges as opposed to bytes for better parlance of media interfaces.

And then we have a didFinishDownloadingTo location delegate method which informs you where the download gets downloaded to.

So this is similar to URLSessionDownloadDelegate's protocol method here.

But your expectation for what you need to do when you get this delegate method is a little bit differently.

So I'll point that out in a second.

So here's an example of our progress delegate method.

Here, I'm just converting time ranges to a percent complete downloaded.

I'm not going to go into it in detail, but that's how you would do that.

Another thing that might happen during your download is your app might be killed.

The process might be terminated.

And then what do you do?

Your download continues.

Well, to respond to that, you respond exactly how you respond to URLSessionDownloadTasks.

And that's by, when your app launches, you're going to want to set up another background configuration using the same identifier you used to set up your download in the first place and then create a URLSession from that.

And then from there, call the getAllTasks method.

And then here, you can restore your AVAssetDownloadTask which will have the current progress of it.

And you can use this to update any UI that you might have in the app.

And you can even use it to grab the original AVAsset that you used to set up your download.

So once your download finishes, the first thing you're going to want to do is store the location of the download asset.

So this method is called whenever anything is deposited at this location, including when a partially-downloaded content is canceled by your user.

Now you can use this partially-downloaded, cancelled download to resume the download at a future time.

Or you can even use it to playback whatever has been partially there.

But if you really don't want that download there or that content there on your app storage anymore, then here's where you should probably delete it.

So, unlike URLSessionDownloadDelegate, we ask that you do not move the asset from this location.

And we do this because it's really important that the system be able to find these files as the system may want to reclaim disk space in low disk space conditions, and might actually at some point go out and delete this asset from your app container.

So what you're going to want to save is going to be the relative path of the location.

And then from there, you will be able to restore your asset in the future for playback.

So, now that you have a version of your asset downloaded, let's say you want to download additional media selections.

A good place to do this would be in our didCompleteWithError delegate method.

This is the same one that URLSessionTask has.

And the reason why it might be good to do here is because media selections on the same asset are downloaded serially.

And this is generally the right place to do any cleanup of the complete download.

So, to augment with an additional media selection option say, the Spanish audio we would first grab our spanishOption that we want to download.

Then we would mutate the AVMediaSelection that we originally passed in to select the spanishOption from there.

And then once we have that selected, we're going to want to create a new AVAssetDownloadTask preferably on the same session with passing the additional media selection we want to download.

And then we call resume, and then the whole process repeats itself.

And then we could download additional media selection after that if we wanted to.

So let's talk about playing this thing now.

So, say we had a here's an example of how we could create a download task.

And notice that I pass in an asset with a networkURL to it.

When we playback offline HLS, if we still have that original urlAsset available, then we should use that AVAsset instance to play.

And we can grab this from the urlAsset property off the AssetDownloadTask.

If you pass in a new asset with networkURL, AVFoundation will not know where to find the downloaded content from.

So you should use that same asset, and then you will have playback offline.

However, if you no longer have any references to that original AVAsset for example, a long period of time later when your app might not have been running then what you're going to want to do is you're going to want to create a new AVAsset from the download location that we supplied to you in the didFinishDownloadingTo delegate method and create a PlayerItem with that.

Now even in this case, if you want to do an operation on the asset for example, augment with an additional media selection download you really should reuse that same asset.

You shouldn't have a bunch of AVAssets pointing toward the same fileURL on disk.

This helps AVFoundation be efficient about reusing already-downloaded media data.

So we also have an additional class to help you be stringent about what you let your users play offline, and this is called AVAssetCache.

You grab this from the asset cache property on an AVAsset.

And it can do two things.

One is that it can tell you if any rendition of this asset is playable offline.

And the other thing it can do is it can tell you whether a specific media selection option is available offline.

Now, remember, if it's not available offline, it's still playable so long as you're connected to the network and that media selection option is still available on the server that you downloaded it from.

So, that's how you use AVAssetDownloadTask and play the subsequent downloaded asset.

And I should note that number five here, playing the download asset disk, that can happen really at any point in this process.

It can even happen before you start the download, or even during the download.

And AVFoundation will be efficient about trying to reuse what has already been downloaded from the network.

So I want to switch gears a little bit here and talk about securing the offline content.

So, last year, we introduced FairPlay Streaming to bring strong content protection to the HLS ecosystem.

However, in iOS 9, FairPlay Streaming requires a live connection to a key server, which doesn't really work for offline playback.

So, this year, we've extended FairPlay Streaming to work without an internet connection for the offline HLS case.

Now we do this by packaging your keys in a manner that is safe to store on disk and reuse for offline playback.

But your app is still expected to store this key itself and respond to key requests from AVFoundation on every single playback without hitting the network once it has its key cached.

So support for these keys does require a change in your key servers, specifically a new TLV value so your key servers have to explicitly opt-in keys to be eligible for offline playback.

And finally, we have a caveat for FPS Keys.

They must be declared in your master playlist as session keys if you want to make sure that they are downloaded.

So, storing keys for offline FairPlay Streaming really builds off the request flow of online FairPlay Streaming.

And so what you have to change really just has to do with how you modify this request flow.

I'll give a quick recap of it here to point out what you need to do.

But for more information, you should watch last year's talk where we go into this in a lot of detail.

So just like in FairPlay Streaming for playback, AVFoundation will download your playlist and trigger a key request if it finds a key tag, which will give your app a chance to save the key for offline playback.

So the first thing your app would do is it would call back to AVFoundation to get a streaming key request.

Then it would talk to your server, which would take that streaming key request and give you a content key context or a CKC.

And then in online FairPlay Streaming, this is where you would finish a request flow and simply return to AVFoundation with that CKC.

However, in offline FairPlay Streaming, here's where you need to give that CKC to AVFoundation so AVFoundation can freeze-dry it into a format that is safe for you to store into your app storage.

And this is something that you can also reuse on subsequent playbacks.

So once you have this thing written into your app storage, you want to return that freeze-dried key to AVFoundation to complete the request.

Now if your app were offline, then all you would have to do is read that freeze-dried key from your app storage and return it to AVFoundation so you could do this without touching the network.

So let's look at how this actually changes your code.

So in online FairPlay Streaming, these key requests happen as a part of AVAssetResourceLoader, specifically as a delegate method on AVAssetResourceLoadingRequest.

So the first thing you would do once you get one of these requests is ask AVFoundation to create a streaming content key request or an SPC, which sends that to your server.

And then you get a CKC back, and then you respond to AVFoundation with that.

Now in offline FairPlay Streaming, we have a new method.

This is the method that you use to freeze-dry the key.

You pass in the CKC you get back from the server, and it returns back that data blob that is saved for you to store offline.

We also have a new content type as part of this request flow the persistentContentKey type and a new option to pass into our streaming content key request data for app method.

So here is how we would modify that key request flow for a key that we want to save offline.

The first thing that we want to do is ask AVFoundation to create an SPC.

The difference here is that we need to pass in the required persistentKey option.

So that way, AVFoundation and your key server knows to request a key that will be eligible for that freeze-dry process.

So it sends this SPC to your server to get the CKC.

And then once you get the CKC back, you call your persistentContentKey method to create this freeze-dried thing that you can save to disk.

Then, you want to actually write that thing to your disk.

And then, you want to set your content type to a persistentContentKey type and then finish the downloading request with that persistent CKC that you saved to your disk.

So now if you are offline, or say you already satisfied this key request previously, then this is what you will do.

You get your resource loadingRequest and you just simply read your freeze-dried key from your app storage.

Then you set your persistentContentKey type, and you respond to the key loadingRequest all without hitting the network, which is obviously required as you may potentially be in airplane mode and don't have any network connectivity during this process.

So that's how you modify FairPlay Streaming to work with your offline HLS content.

So now I want to talk about best practices for managing your assets.

So it's important to note that these downloads do contribute to your app's disk usage, so you want to be a little bit mindful here.

You should really clean up any unneeded assets on disk.

Specifically, you should definitely provide a way to your users to see everything that you have downloaded and allow them to delete things that they don't want any more.

Remember that downloads that users cancel do remain on disk.

Now you can use those to resume downloads in the future at some point if you like.

But if that's not what you want, then you should take care to go ahead and delete those assets once they are cancelled.

We're asking that you keep downloads driven by explicit user actions.

So you shouldn't just go ahead and download a bunch of stuff that your user might not actually want offline.

We've gone ahead and taken the liberty of opting out these downloads from iCloud backup.

As I mentioned earlier, you should be prepared for the system to delete your assets to reclaim disk space in low disk space conditions.

Now this won't happen while your app is running, ever.

This will only happen while your app is terminated.

After your app is launched, you should be a little bit defensive about assuming whether you actually have the assets that you had previously downloaded on disk.

You do not move your assets from the location which we provide them to.

Again, it is very important that the system be able to find these assets.

And finally, because these downloaded assets may contain media selections that have not be downloaded, be careful not to mutate the asset that you have on your servers.

If you really must change, say, like the location of a media playlist and stuff like that, and mutate your master playlist, then you should just host that modified asset at a new URL rather than changing the one that's already there.

So, that's offline HLS.

And we're really happy to bring this to you and excited to see you adopt it.

So, in summary, we have new features for you to use this year.

We have our fragmented MP4 support which will bring a common media format across different platforms that you may be supporting.

Remember that this is compatible with all HLS features and requires minimal changes to your HLS playlist to adopt.

Next, we have our new in-playlist metadata using the DATERANGE tag.

and this is just really good to use metadata with any live content that might have dynamically updating metadata, like ad boundaries and stuff like that.

And finally, we're giving you an offline HLS playback.

So part of this we have a powerful media downloading engine as well that you can configure media selections and quality to download.

And we're also extending our FairPlay Streaming support to work without an Internet connection.

So for more information, including some sample code on how to use AVAssetDownloadTask and more documentation on these new editions, go see our page on the development site for this session.

I want to point out the session that Roger mentioned earlier, our version of authoring and validation talk.

You can watch this video in the WWDC app right now.

I highly recommend watching it.

So, thanks for coming.

And I hope you enjoy the rest of the conference.

[ Applause ]

Apple, Inc. AAPL
1 Infinite Loop Cupertino CA 95014 US