iQiyi Video Buffering Analysis and Handling Experience

Video buffering (卡顿) is a common issue when users watch on-demand videos. For a leading video platform like iQiyi with over 200 million daily active users, noticeable buffering occurs frequently among user groups.

Buffering ratio is defined as the number of users experiencing at least one buffering event in a 5‑minute window divided by the total independent users in that window. Changes in statistical dimensions (province, carrier, IDC, client type) affect both numerator and denominator accordingly.

To ensure smooth viewing, iQiyi relies on operations personnel to handle buffering anomalies. The platform maintains an average buffering ratio below 2% year‑round, attributing this to efficient fault‑handling processes and strict management regimes.

Based on years of experience, iQiyi categorizes buffering anomalies into four major types: internal system problems (≈10%), server‑side problems (≈30%), operator network problems (≈40%), and user‑end problems (≈20%).

Internal system issues involve client network download modules, playback event statistics modules, video CDN scheduling systems, and video CDN servers (Nginx). Problems in these areas may stem from strategy changes, statistical rule updates, scheduler latency, or server faults such as IDC traffic saturation, switch failures, or single‑server malfunctions.

Operator network problems include uplink link issues, provincial backbone or inter‑province faults, and other routing or hijacking problems. Detection often relies on tracing user IP segments and coordinating with data‑center staff; resolution efficiency is limited by operator response times.

User‑end problems arise from edge‑device faults, malicious users, or special behaviors of popular‑drama audiences. iQiyi employs a user‑portrait system to analyze buffering users in real time and feeds results back for alerting, enabling rapid identification of anomalous user behavior.

To detect anomalies, iQiyi built a “buffer perception system” that analyzes fine‑grained buffering ratios, compares with historical data, filters noise, and uses clustering to output a suspicion probability. The system also combines client buffering ratios with server‑side download speeds, employing SVM to predict expected buffering from download speed; large discrepancies indicate statistical system errors.

Internal management mechanisms require any production system change to notify the buffering‑responsibility department and describe its impact. Bugs causing buffering anomalies (statistical or real) incur responsibility, and unresolved buffering faults beyond 24 hours (except operator issues) also trigger accountability.

Monitoring automation includes multi‑dimensional traffic monitoring, Fourier‑transform‑based frequency‑domain analysis to boost switch‑traffic alarm accuracy to >97.6%, and real‑time link monitoring.

Link‑level monitoring leverages iQiyi’s product lines to deploy wide‑coverage hardware probes, achieving nationwide real‑time service monitoring without affecting user experience. This system has yielded over ten patent applications.

User research integrates buffering data into a behavior‑analysis system, treating buffering as a dimension for model training. Offline analysis of buffered‑user data enables automatic cause identification within ~3 minutes per query.

The combined “buffering exception automatic analysis system” connects monitoring alerts to responsible personnel and performs offline queries to deduce the most likely cause of a buffering incident.

As a result, iQiyi resolves over 93 % of non‑operator‑related large‑scale buffering anomalies within 15 minutes and >99.9 % within 24 hours. Operator‑caused anomalies are located and complained about within 20 minutes in >97 % of cases.