blacklist the playlist that has stopped being updated and never blacklist the final playlist available

[zhuangs]

Description

Detect that playlist has stopped being updated in a timely manner and blacklist the playlist
Never blacklist the final playlist available #1026

[mjneil]

I think it would be good to pull out this section of calculations to a separate function as this event handler is getting quite gnarly

[zhuangs]

changed this

[mjneil]

seekableEnd, lastBufferedEnd and endTime have the potential of being NaN at this point since seekableEnd and lastBufferedEnd may never be assigned a value if the conditionals for the if statements fail, and playEnd may be null from this.playlistEnd() affecting the calculation of endTime

[zhuangs]

I thought seekableEnd, lastBufferedEnd, endTime being NaN may only happen at the very beginning of a playlist and that isn't the case we should blacklistCurrentPlaylist, is this correct? In that case, it doesn't affect the if statement. Did you mean we should check seekableEnd, lastBufferedEnd, endTime isNaN() inside the if?

[mjneil]

With lastBufferedEnd for example, you declare lastBufferedEnd at the start of this function, so it is undefined. The only location lastBufferedEnd is assigned a value is on line 378, which is inside the if (buffered.length) check. If buffered.length === 0 then lastBufferedEnd will still be undefined and any calculation involving lastBufferedEnd e.g. endTime = playEnd - lastBufferedEnd will result in NaN.

With that said, looking at this again, it may be ok for this to happen. Since any comparison with NaN will return false, seekable.length === 0 or buffered.length === 0 probably means we don't have enough information to know if the playlist stopped updating and probably shouldn't blacklist it. This will require testing to see if we can create a scenario in which those values result in NaN but we still want to blacklist

[mjneil]

So I was giving this more thought and I think we can avoid worrying about it all together by calling this method and blacklisting the playlist outside of the loadedplaylist listener. Playlist Loader already knows when a playlist it has loaded didn't actually update the live window. The isPlaylistOutdated_ method more so checks if the player is stuck at the end despite loading "fresh" playlists. If instead playlist loader triggered an event at that moment it knows the playlist isnt a fresh update, then MasterPlaylistController can listen to that and run the checks for playlistOutdated and if those checks pass, then we blacklist. This way we wont run these checks every time the playlist is refreshing correctly and run the risk of accidentally blacklisting a playlist we don't want to

[zhuangs]

changed as suggested

[mjneil]

It isn't necessary to attach this function to MasterPlaylistController, you can just call playlistEnd directly where you need it. On top of that, if you attach playlistEnd to the default export of playlist.js, the default export is already included in the Hls object available in this file. If you go that route, you do not need to import this function in a seperate import call, and instead could just call Hls.Playlist.playlistEnd()

[zhuangs]

I attached this function to MasterPlaylistController because I was trying to call this function in the videojs-contrib-hls.test.js file. but maybe there is better way to do this

[mjneil]

Ah didn't notice that. Hmmmm, its not a big deal to attach the function, but if you come up with another solution that doesnt require it, thatd be great

[zhuangs]

yes, I found directly override Hls.Playlist.playlistEnd() in the test file is better :)

[mjneil]

this function is very similar to seekable defined just below. Is there any way we can reduce the amount of repeat code here?

[zhuangs]

changed this

[mjneil]

Would be good to have a some tests for playlistEnd

[mjneil]

This should return null by default isntead of 0. If there isn't enough information to determine expired time, returning 0 would be sending false information if there actually is expired time.

[zhuangs]

Did this in the first place but then I found there is a case that the playlist contains the endList, so the seekableEnd will return the duration of playlist but the seekableStart returns null. I think that's why I changed here to return 0 and check for null on the sync points instead of checking expired.

[mjneil]

That is a good point, but I don't think returning 0 is the right call here since there are live scenarios where we may not have any sync points, but there has been expired time, we just can't calculate it. Since endList indicates a VOD asset, we actually already have a known sync point then just off knowing endList alone, which would be { time: 0, mediaSequence: 0 } since VOD always starts at time 0 segment 0. I'd say it would make most sense to update getPlaylistSyncPoints to set expiredSync to { time: 0, mediaSequence: 0 } if playlist.endList, otherwise set to playlist.syncInfo, and finally resorting to null. That way calculateExpiredTime can properly calculate for VOD case, while still returning null in the live case where we don't have enough information

[zhuangs]

fixed this

[mjneil]

Since you moved getPlaylistSyncPoints to calculateExpiredTime calling it again here just to check for null on expiredSync segmentSync is unnecessary work. If calculateExpiredTime starts returning null when it can't calculate expired, you can just check for null on expired instead of checking the sync points

[zhuangs]

fixed this

[mjneil]

minor, but I think we can come up with a better name for the method. Maybe isPlaylistOutdated_, isPlaylistStagnant_, isPlaylistUpdating or something along these lines that is a bit more descriptive of what the method does. (note the return value should reflect the name logically). This method should also be private, indicated by the _ appended to the end of the name

[zhuangs]

fixed this

[mjneil]

maybe name the event playlistunchanged ?

[zhuangs]

renamed!

[mjneil]

I know I've suggested renaming this already, but now that we have an event that signifies the playlist is outdated/not refreshing, and this method more checks if we are stuck at the end of the playlist, so maybe somthing like stuckAtPlaylistEnd_or isAtPlaylistEnd_. Then our logic sort of semantically reads like if the refreshed playlist is unchanged and the player is at the end of the playlist, blacklist this playlist.

[zhuangs]

makes sense! renamed!

[mjneil]

I think it would be good to add a check between the two refreshes to make sure the playlist wasnt blacklisted early. (you may have to move the play and playing triggers up, not sure though)

[gesinger]

Would be good to have some tests for stuckAtPlaylistEnd_ directly.

[gesinger]

Would be good to have some tests for stuckAtPlaylistEnd_ directly.

[gesinger]

extra comment

[gesinger]

Can remove a lot of lines up top if we create the variables where they are first used.

[gesinger]

If there is no seekable range, can we be technically stuck at playlist end?

[gesinger]

Since this is only used in one place, it may be clearer to just use it in the calculation directly: endTime = Hls.Playlist.playlistEnd(playlist) - lastBufferedEnd; . We should also change the comment to say it is using the absolute end of the playlist, and not the safe live end.

[gesinger]

While NaN can sometimes be OK, it is usually good to try to avoid using it. In this case, where we need the amount of time buffered, this should be 0 instead of NaN.

[gesinger]

If this function just calls the other function, do we need both?

[gesinger]

Would be good to keep some documentation about the method, particularly the definition of useSafeLiveEnd (i.e., "live playlists should not expose three segment durations worth...")

[gesinger]

Can just return directly instead of creating a variable for it.

[gesinger]

We might be able to use standardXHRResponse to save some lines. We can also use this.requests.shift to make it so that the latest outstanding request is always at the beginning of the array (no need to keep track of them via indexing).

[gesinger]

playlist wasn't blacklisted

[gesinger]

Indentation

[zhuangs]

changed! sorry for keeping making this mistake

[mjneil]

Since stuckAtPlaylistEnd_ only uses seekable to check if we have a seekable range, but doesn't use seekable.end(0) for any of its checks, is there even a point to checking not stuck at playlist end when currentTime not at seekable end even if the buffer is empty?

If we still want to keep this part in, I'm not sure having playlistEnd and seekable.end(0) both be the same (130) is representative of an actual scenario since seekable.end is the time since playback started to minus 3 segment durations from the end of the window, where playlistEnd is time since playback started to minus 0 segment durations from the end of the window.

[gesinger]

Good point. I think that leaving in tests that make use of seekable values is good, even if stuckAtPlaylistEnd_ doesn't use them, since the tests shouldn't know too much about the internal workings of the function. But you're right that the scenario of playlistEnd equaling seekable's end shouldn't be a common case. However, in the event that a playlist changes from LIVE to VOD (at the end of a stream), and the only update was the addition of #EXT-X-ENDLIST, seekable's end should equal playlistEnd, and I think it's possible that this case is a valid one.

[mjneil]

When a live stream is finished, usually an EXT-X-ENDLIST tag is applied to the end of the playlist even though it wasn't originally a VOD. This would set expiredSync to { time: 0, mediaSequence: 0 } even if that isn't an actual sync point. We should prioritize playlist.syncInfo if it exists, otherwise { time: 0, mediaSequence: 0 } if playlist.endList is true, otherwise null.