How Tencent Scales NBA Live Streams to Millions: Behind the Tech and Operations

Preface

We gradually improved our technology while tackling major challenges, and this article shares over a year of work on massive live streaming, focusing on the representative case of 1,300 NBA broadcasts per year.

1. The Rapid Growth of Live Streaming

Since 2015, live streaming has surged due to four drivers:

Technology drive Advances in smart hardware, smartphones, and network bandwidth make live viewing more convenient.

Commercial drive Content IPs such as sports and concerts generate massive live events, enriching the content ecosystem.

Content ecosystem Monetization methods like memberships and tipping energize the industry.

Mass demand Audiences crave real‑time updates and immersive experiences, driving high engagement.

These factors, combined with technical support, have propelled live streaming to new heights.

2. Characteristics of an Excellent Live Stream

Key technical requirements include crystal‑clear picture quality (e.g., 1080p must look better than 720p), audio‑video synchronization, low latency, and multi‑angle coverage. Failure in any of these areas leads to poor user experience and brand damage.

3. Technical Choices for Live Streaming

3.1 Video Live Stream Process

The workflow spans video capture, transmission, packaging, encoding, pushing, transcoding, distribution, decoding, and playback, involving up to eighteen stages and multiple camera angles.

3.2 Playback Experience Challenges

We identified four main issues: long‑distance transmission (≈18,000 km from the US to China), content localization and packaging, multi‑angle and multi‑resolution delivery, and ensuring smooth playback without stutter or artifacts.

4. Transmission Challenges

4.1 Transmission Losses and Stalls

Long‑haul UDP transmission can cause packet loss, leading to visual glitches. We introduced error‑correction matrices (e.g., 10×10) to recover lost packets, reducing visible artifacts to a fraction of a percent.

4.2 Costly Satellite Alternatives

Satellite links are prohibitively expensive (≈50× network transmission cost), making them unsuitable for 1,300 annual events.

4.3 Optimization Solutions

Fault‑tolerant coding Adding redundant check packets allows reconstruction of lost data, improving reliability to 0.1% loss.

Multi‑link backup Deploying three redundant network paths (red, green, yellow) ensures continuity even if one link fails.

5. Production Technology Challenges

5.1 Visual Optimization – Subtitles

We add Chinese subtitles to convey player information during broadcasts.

5.2 Visual Optimization – AR

AR overlays enrich the viewing experience with interactive data.

5.4 Visual Optimization – Multi‑Angle

Providing multiple camera angles (e.g., left, right, under‑basket) enhances engagement.

6. Playback Challenges

6.1 Playback Smoothness

Users expect the video to start within two seconds; each additional second increases abandonment risk.

Even a one‑second stall reduces viewing time and raises churn.

6.2 CDN Solution

We operate 500 CDN nodes across China (including remote regions) and use IP‑based routing to minimize latency.

6.3 Massive User Surge Management

During NBA starts, peak concurrent users can exceed 2 million per minute, requiring robust load handling.

6.4 Scheduling Strategies

SNMP data collection Directly gathering switch traffic reduces monitoring latency.

Predictive traffic shaping Analyzing entry patterns allows pre‑emptive traffic diversion before a data center becomes overloaded.

6.5 Faster First‑Frame Rendering

Prioritizing I‑frame delivery reduces the time to first visible frame by ~200 ms, improving perceived start‑up speed.

7. Massive Monitoring Challenges

7.1 Monitoring Goals

We monitor low‑level metrics (CPU, memory, network) and high‑level playback indicators (startup time, stall rate, bitrate).

7.2 Log Analysis Bottleneck

Daily log volume reaches 2 trillion entries, requiring minute‑level analysis rather than hours‑level.

7.3 Big‑Data Processing Solution

Data from all client platforms is collected, cleaned, sent via Kafka to Spark clusters for real‑time aggregation, producing dashboards and alerts that handle >2000 billion records and >100 TB of data.

8. Conclusion

Massive live‑streaming operations demand a comprehensive technical stack—from network transmission and CDN to monitoring and big‑data analytics—to deliver reliable, low‑latency experiences for millions of users.