Maximizing Viewer Engagement with Automated Video Streaming Diagnostics

Table of Contents

1. Introduction
2. The Need for Auto Tech Diagnostics Alert
3. The Growth of Internet Video Streaming
4. The Complexity of the Streaming Pipeline
5. Detecting Anomalies in Video Metrics
6. Constructing a Diagnose Graph
7. Running the Detection Algorithm
8. Identifying the Root Cause
9. Machine Learning for Small Groups
10. Distributed Algorithm with Spark
11. Performance in Production Environment
12. Future Improvements and Challenges

Introduction

Welcome to our presentation on Auto Tech Diagnostics Alert. In this article, we will discuss the project built by the engineering team at Kuvira, which focuses on detecting anomalies in the video streaming pipeline and automatically diagnosing the root cause. We will dive into the details of the product and its significance in the rapidly growing industry of online video streaming. The article aims to provide a comprehensive understanding of the system's functionality and the need for such advanced diagnostics in ensuring optimal viewer engagement.

The Need for Auto Tech Diagnostics Alert

The tremendous growth of internet video streaming has presented numerous challenges for content publishers. With the shift towards online streaming, it has become increasingly difficult to guarantee a seamless end-to-end streaming experience. The complex streaming pipeline, comprising multiple entities and paths, introduces the risk of silent failures that degrade video quality. For example, a single bug in a video player can result in buffering issues for viewers using specific devices, leading to a decline in viewer engagement and increased churn rate.

To address this challenge, a system like Auto Tech Diagnostics Alert is required. This system aims to provide near real-time detection of quality issues along the streaming pipeline, allowing content publishers to identify and resolve the root cause of these issues promptly. By monitoring and analyzing key performance indicators (KPIs) at various entities in the pipeline, the system can pinpoint anomalies and initiate appropriate actions, ultimately maximizing viewer engagement and ensuring the success of content publishers.

The Growth of Internet Video Streaming

In recent years, the industry of online video streaming, also known as OTT (Over-The-Top), has experienced rapid growth. With the increasing popularity of streaming platforms, more and more content publishers have transitioned from traditional cable television to online streaming. In fact, it is estimated that by 2019, over 80% of global internet traffic will be dedicated to online video streaming.

This exponential growth presents both opportunities and challenges for content publishers. While online streaming offers a broader reach and flexibility, ensuring a successful end-to-end streaming experience remains a significant concern. The complexity of the streaming pipeline, as well as the multitude of devices and networks involved, makes it crucial for content publishers to have a reliable system in place to monitor, diagnose, and resolve issues that may arise, ultimately ensuring a seamless viewer experience.

The Complexity of the Streaming Pipeline

The streaming pipeline comprises various entities, each playing a vital role in delivering video content to viewers. However, any entity along the path can fail silently, resulting in degradation of the overall video quality. To illustrate the complexity of the streaming pipeline, imagine a scenario where viewers can choose different devices to stream a live event. Each device can stream from multiple CDNs (Content Delivery Networks), and these CDNs rely on live encoders. An issue with a single live encoder can cause one CDN to stop receiving data from the origins, cascading into playback failures and buffering issues for viewers streaming from that CDN.

Identifying the root cause of such issues becomes a challenging task. Without a comprehensive system in place, content publishers may struggle to determine which part of the pipeline requires attention and resolution. This lack of visibility hampers their ability to take effective actions, resulting in continued viewer suffering and increased churn rate. Therefore, it is crucial to have a system like Auto Tech Diagnostics Alert that can systematically detect anomalies, diagnose the root cause, and provide actionable insights to address streaming issues promptly.

Detecting Anomalies in Video Metrics

To detect anomalies in the streaming pipeline, the Auto Tech Diagnostics Alert system employs a two-step approach. The first step involves estimating a baseline for the quality metrics based on historical data. This baseline serves as a reference point to determine whether a quality metric is within the normal range or exhibits abnormal behavior. The system then calculates a tolerance threshold, which represents several standard deviations above the baseline. Any metric value below the threshold is considered normal, while values above the threshold indicate anomalies.

In addition to detecting anomalies, the system also focuses on quantifying the impact of these anomalies. By computing the area of each spike in the time series data, the system can measure the severity of the issue. A spike with a larger area implies a more significant impact on video quality. If the computed area exceeds a predefined threshold, the system flags the spike as an anomaly, signaling the need for further investigation and resolution.

Constructing a Diagnose Graph

To efficiently identify the root cause of anomalies, the system constructs a diagnose graph. This graph represents the parent-child relationships between groups of video sessions based on various dimensions such as device, CDN, and video publisher. For example, a group representing sessions from iPhones can be split into subgroups corresponding to different CDNs. This hierarchical structure allows the system to systematically analyze and trace anomalies across different entities in the streaming pipeline.

Running the Detection Algorithm

Once the diagnose graph is constructed, the system runs the detection algorithm for each group defined in the graph. This algorithm compares the time series data of each group against the baseline and tolerance thresholds to determine whether an anomaly is present. By marking anomalies at each level of the graph, the system can identify groups with abnormal behavior and proceed with the root cause analysis.

Identifying the Root Cause

To identify the root cause of anomalies, the system employs a top-down search approach in the diagnose graph. Starting from the top level, the system recursively drills down to find the root cause group for each anomaly. The search ends when a group is found with no subgroups exhibiting abnormal behavior. This group is then identified as the root cause of the issue. By systematically tracing anomalies and identifying the root cause, content publishers can allocate resources more effectively and fix issues that impact multiple entities in the streaming pipeline.

Machine Learning for Small Groups

In cases where small groups have limited traffic and unique time series patterns, the system utilizes a machine learning module to estimate their performance. The module is trained on data from other video sessions and provides predictions of average performance for small groups at each minute. These predictions, along with the computed time series data, are then subjected to the same anomaly detection algorithm. This approach ensures that even smaller groups receive accurate anomaly detection and diagnosis, contributing to a more comprehensive and reliable system.

Distributed Algorithm with Spark

To handle the scale and complexity of processing video sessions data, the Auto Tech Diagnostics Alert system leverages the Spark framework for distributed computing. The system loads the sessions data into Spark executors, shuffles the groups based on video publisher IDs, and constructs diagnose graphs for each executor. The detection and diagnosis algorithms are then applied to each graph, and the root cause is stored in a database for further analysis and action. This distributed approach allows the system to efficiently process a large volume of video sessions and provide timely alerts to content publishers.

Performance in Production Environment

The performance of the Auto Tech Diagnostics Alert system has been remarkable in a real production environment. With the ability to support 25 video publishers and handle a substantial number of sessions per minute, the system has successfully identified CDN failures and provided actionable insights to content publishers. The time series data visualization and session details have enabled content publishers to debug issues promptly and improve viewer engagement. However, continuous evaluation of accuracy and latency optimization remains a priority for future enhancements.

Future Improvements and Challenges

Looking ahead, the Auto Tech Diagnostics Alert system aims to refine its detection and diagnosis algorithms further. Evaluating accuracy and soliciting customer feedback will play a crucial role in improving the system's performance. Additionally, addressing the challenge of defining impact thresholds and refining the anomaly detection criteria will contribute to more precise and actionable alerts. As the industry of online video streaming continues to evolve, the system will adapt and accommodate emerging technologies and complexities, ensuring that content publishers can deliver an exceptional streaming experience to their viewers.

👉 Highlights

• Auto Tech Diagnostics Alert system detects anomalies in the video streaming pipeline and provides automated root cause diagnosis.
• The growth of internet video streaming necessitates advanced diagnostics to ensure viewer engagement.
• The complex streaming pipeline presents challenges in identifying and resolving issues promptly.
• Anomaly detection involves estimating baselines, calculating tolerance thresholds, and quantifying the impact of anomalies.
• The construction of a diagnose graph enables systematic root cause analysis across different entities in the streaming pipeline.
• The detection algorithm, machine learning for small groups, and distributed processing with Spark are key components of the system.
• The performance of the system in a real production environment has been highly successful in identifying and resolving issues.
• Continuous improvement and customer feedback are essential in enhancing accuracy and optimizing latency.

FAQ

Q: What is Auto Tech Diagnostics Alert? A: Auto Tech Diagnostics Alert is a system developed by the engineering team at Kuvira to detect anomalies in the video streaming pipeline and provide automated root cause diagnosis.

Q: Why is it essential to detect anomalies in the streaming pipeline? A: Detecting anomalies is crucial to ensure a seamless end-to-end streaming experience for viewers. Anomalies can lead to buffering issues, playback failures, and decreased viewer engagement, which, in turn, can result in increased churn rate for content publishers.

Q: How does the system identify the root cause of anomalies? A: The system uses a hierarchical diagnose graph to trace anomalies across different entities in the streaming pipeline. By systematically drilling down and analyzing groups, the root cause of the issue can be identified.

Q: What role does Spark play in the system? A: Spark is utilized for distributed computing, enabling efficient processing of a large volume of video sessions data. It allows for the construction of diagnose graphs, running the detection and diagnosis algorithms, and storing the results for further analysis.

Q: How can content publishers benefit from the Auto Tech Diagnostics Alert system? A: The system provides actionable insights to content publishers, allowing them to promptly identify and resolve issues impacting the viewer experience. This leads to improved viewer engagement and reduced churn rate.

🌐 Resources:

• Kuvira