Week3 - Using Machine Learning to Improve Streaming Quality at Netflix

[kymr]

Title

• Using Machine Learning to Improve Streaming Quality at Netflix

Summary

• Providing a quality streaming experience for this global audience is an immense technical challenge
• install and maintain servers throughout the world
• algorithms for streaming content from those servers to our subscribers’ devices
• diverse viewing behavior
  • Viewing/browsing behavior on mobile devices is different than on Smart TVs
  • Cellular networks may be more volatile and unstable than fixed broadband networks
  • Networks in some markets may experience higher degrees of congestion
  • device groups have different capabilities and fidelities of internet connection due to hardware differences
• Network quality characterization and prediction
  • the average bandwidth and round trip time supported by a network are well-known indicators
  • stability and predictability make a big difference when it comes to video streaming
  • A richer characterization of network quality would prove useful for analyzing network, determining initial video quality and/or adapting video quality throughout playback
  • How can we incorporate longer-term historical information about the network and device?
  • combine temporal pattern recognition with various contextual indicators to make more accurate predictions of network quality
• Video quality adaptation during playback
  • Adaptive streaming algorithms are responsible for adapting which video quality is streamed throughout playback based on the current network and device conditions
  • The quality of experience can be measured in several ways
  • including the initial amount of time spent waiting for video to play
  • the overall video quality experienced by the user
  • the number of times playback paused to load more video into the buffer (rebuffer)
  • the amount of perceptible fluctuation in quality during playback
  • These metrics can trade off with one another
  • The feedback signal of a given decision is delayed and sparse
  • when learning optimal control algorithms, and machine learning techniques (e.g., recent advances in reinforcement learning) have great potential to tackle these issues.
• Predictive caching
  • statistical models can improve the streaming experience is by predicting what a user will play in order to cache (part of) it on the device before the user hits play
  • By combining various aspects of their viewing history together with recent user interactions and other contextual variables
• Device anomaly detection
  • Netflix operates on over a thousand different types of devices
  • it is not uncommon to cause a problem for the user experience
  • the app will not start up properly, or playback will be inhibited or degraded in some way
  • Detecting these changes is a challenging and manually intensive process.
  • Alerting frameworks are a useful tool for surfacing potential issues but oftentimes it is tricky to determine the right criteria for labeling something as an actual problem
  • Fortunately, we have history on alerts that were triggered as well as the ultimate determination (made by a human) of whether or not the issue was in fact real and actionable
  • it is often challenging to determine the root cause
  • Statistical modeling can also help us determine root cause by controlling for various covariates
• statistical modeling and machine learning methods can improve the state of the art
  • there is sufficient data
  • the data is high-dimensional and it is difficult to hand-craft the minimal set of informative variables for a particular problem
  • there is rich structure inherent in the data due to complex underlying phenomena

Reference

• https://medium.com/netflix-techblog/using-machine-learning-to-improve-streaming-quality-at-netflix-9651263ef09f

[kymr]

Words

• One of the common questions we get asked
• given the recent hype around
• solution in search of a problem
• some of the technical challenges we face
• live outside the United States
• is an immense technical challenge
• engineering effort required to
• as well as
• As we expand rapidly
• with widely varying capabilities
• fixed broadband networks
• higher degrees of congestion
• different capabilities and fidelities
• fluctuating conditions
• aspects of the user experience
• are well-known indicators of network quality
• when it comes to video streaming
• throughout playback
• during real viewing sessions
• what throughput will look like
• How can we incorporate longer-term historical information
• Even if we cannot predict exactly
• a microwave turning on
• going through a tunnel
• characterize the distribution of throughput
• Since we are observing these traces at scale
• temporal pattern recognition
• various contextual indicators
• to support different network and device capabilities
• The figure below illustrates
• can be measured in several ways
• the initial amount of time spent waiting for video to play
• the number of times playback paused
• the amount of perceptible fluctuation in quality
• is divided up into
• can trade off with one another
• increase the risk of a rebuffer
• download more video up front
• at the cost of increased wait time
• The feedback signal of
• an aggressive switch to higher quality may not have immediate repercussions
• could gradually deplete the buffer
• credit assignment” problem
• Another area in which
• by predicting what a user will play in order to cache it
• we can exploit the fact that
• is very likely to play
• By combining various aspects
• the model’s likelihood of caching
• what the user actually ended up playing,
• while respecting constraints around resource usage
• We have seen substantial reductions in the time spent waiting for video to start
• ranging from laptops to tablets
• New devices are constantly entering into
• existing devices often undergo updates
• These often go without a hitch
• playback will be inhibited or degraded in some way
• gradual trends
• that can accumulate over time
• may slowly degrade performance on a particular device
• it was not immediately noticeable
• manually intensive process.
• surfacing potential issues
• oftentimes
• A “liberal” trigger will end up with too many false positives
• resulting in a large amount of unnecessary manual investigation
• whereas a very strict trigger may miss out
• we have history on alerts that were triggered
• whether or not the issue was in fact real and actionable
• predict the likelihood
• a given set of measured conditions constitutes a real problem.
• Even when we’re confident
• we’re observing a problematic issue
• Was it due to a fluctuation in network quality
• change that was rolled out
• A firmware update issued by the device manufacturer?
• Is the change localized to a particular device group
• by controlling for various covariates.
• we’ve already seen large reductions in overall alert volume
• while maintaining an acceptably low false negative rate
• which we expect to drive substantial efficiency gains
• The aforementioned problems
• there is sufficient data
• it is difficult to hand-craft
• the minimal set of informative variables
• there is rich structure inherent in the data
• due to complex underlying phenomena
• as we stream video under increasingly diverse network and device conditions
• If these problems excite you