get_best_segments can sometimes take a very long time,
we don't want to stop other work from happening while it's ongoing.
So we ask gevent to run other things until there's no other work to do,
then we do one hour, then check back with gevent again.
In combination with the performance improvements, this should mean we don't block
other things from running for more than a few hundred ms at most.
strptime is much faster but can't handle as varied formats.
But in this case we fully control the format, so there's no reason not to use it.
Profiling suggests we spend about 80% of our time in get_best_segments just parsing dates,
so this is a signifigant performance gain.
The prometheus client uses a threading.Lock() to prevent shared access to
certain metric state. This lock is taken as part of doing collection, as well
as during metric.labels().
We hit a deadlock where our stack sampler signal arrived during a collection,
when the lock was held. This meant that flamegraph.labels() blocked forever,
and the lock was never released, hanging all metrics collection.
Our solution is a hack, which is to reach into the internals of our metric object
and replace its lock with a dummy one. This is reasonably safe, but only as long as
the prometheus_client internal structure doesn't change signfigiantly.
To preserve independence between workers and ensure that a
retry (a worker re-create) actually starts from scratch, we only pool connections
on a per-worker basis.
Furthermore, for the same reason, we only let SegmentGetters use the worker's
pool on their first attempt. After that, they create a new pool to ensure they have a clean retry.
Despite this, the result should be that we're almost always re-using an existing connection
when getting segments or media playlists, unless something goes wrong.
SSL connection setup was measured as almost half the CPU time used by the process,
so this change should result in a signifigant CPU usage reduction.
Backdoor allows the operator to telnet into the given port, and get a python shell
running inside the process, from which you can debug, modify state (eg. set the log level),
or whatever. This is extremely useful for debugging weird states that you encounter randomly
but can't easily reproduce, without restarting the process and needing to wait until it happens again.
This meant that old workers would never shut down, causing us to fetch the same media playlist
and same segments multiple times for no reason, and to never give up in face of (non-403/404) errors
even once we have something else working.
Also fix some logging.
When we're out of touch with twitch for long enough, our segment URL will get
so old that twitch stops returning 403 because our token is expired,
and start returning 404s, presumebly becasue the underlying resource has gone away.
We want to treat these the same.
The function is quite customizable and therefore quite complex, but it allows us to
easily annotate a function to be timed with labels based on input and output,
as well as normalize results based on amount of work done to get a better
picture of the actual amount of time taken per unit of work.
This will help us monitor for performance issues.
Prom client doesn't like you creating two stats with the same name,
even though they have different labels and this makes perfect sense.
I feel like I just need to re-write the prom client at some point - it doesn't actually
do all that much except get in your way, apart from the actual text encoding which I
can steal.
Anyway, in the meantime, we get around this by breaking up metrics into two names,
a "foo_all" and a "foo_ENDPOINT". The foo_all lacks the detailed labels,
but is still labelled by endpoint and can be used more easily.
The foo_ENDPOINT labels have more information but require messier PromQL as you need to
match on a name regex if you want to look at more than one specific endpoint.
I had to go to some effort to get nice labelling,
which also meant none of the existing libs for this were any good,
but this works well enough.
Exposes the metrics on /metrics.
The calculations were backwards, so instead of cutting a video by, say, 2 seconds,
it would cut by -2 seconds, which was clamped to 0. So it would never actually cut,
it would always use the closest segment.
Also, once we were actually cutting, we hit an issue where ffmpeg would finish and close
its input early, because we'd reached the end of the cut video, but not all input had been written yet.
This resulted in an EPIPE error (write to closed pipe) in the input feeder. We now ignore that.
This cutter works by only cutting the first and last segments to size,
then concatting them with the other segments, so we only ever process a few seconds
of video instead of the entire video duration.
However, to make this work, care must be taken that the cut segments use the same codecs
as the other segments.
The reason it's experimental is that we are not yet confident in its ability
to cut accurately and without sync issues. We have seen some minor issues when trying to play
back the raw output files, but youtube's re-encoding has consistently smoothed out those issues
and they seem to be highly player-specific. Vigorous testing is needed.
Also note that both methods right now (cat then cut, and cut then cat) only work if all the segments
are cattable, that is they all use the same codecs, have the same resolution, etc.
If a stream were to change its encoding settings, and we were cutting over that change,
both approaches would not work. We should add checks for that scenario (which can only happen
over a stream drop), and if so fallback to a slow method using ffmpeg's concat filter,
which will work even for disparate codecs, though reconciling mismatched resolutions or frame rates
may require further work.
Mike Lang
MasterGunner
Christopher Usher