How Tencent Leverages SkyWalking for Full‑Stack Microservice Observability

Background

In the era of micro‑services, distributed architectures and frameworks make performance analysis and fault location difficult. Tencent Smart Retail needed a way to aggregate real‑time data from all processing stages of its business systems to achieve full‑link performance monitoring.

Goals

The team aimed to build a component platform with core functions such as unified monitoring, call‑chain tracing, and service‑level performance diagnostics.

Why SkyWalking?

After evaluating options, the open‑source SkyWalking project—designed for Docker, Kubernetes, and Mesos environments—matched the requirements. Its architecture consists of three parts: a client (Agent), a collector (Server), and a web UI.

Java Agent probes collect JVM and behavior data.

HTTP and gRPC are used for internal communication.

GraphQL and HTTP render the UI.

Supported storage back‑ends are H2 (for small‑scale debugging) and Elasticsearch.

Service Reporting Practice

Backend services at Tencent use SpringBoot. To avoid code intrusion, the standard start command:

$ java -Dspring.profiles.active=dev -jar target/youshu-app.jar

is extended with the SkyWalking agent by adding the -javaagent flag:

$ java -javaagent:/e/apache-skywalking-apm-bin/agent/skywalking-agent.jar -Dspring.profiles.active=dev -jar target/youshu-app.jar

The UI visualizes topology and trace graphs, as shown in the screenshots.

Tencent Cloud Microservice Observability Platform (TSW)

To reduce operational overhead, Tencent Cloud provides a managed SkyWalking service (TSW). Compared with the open‑source version, TSW offers:

Data Collection (Client) : Flexible reporting, supporting either the TSW probe or the open‑source agent.

Data Processing (Server) : Pulsar Functions smooth traffic spikes; adapters convert data to an OpenTracing‑compatible format; real‑time and offline compute operators handle monitoring, statistics, and long‑term aggregation.

Storage : Supports HBase and HDFS in addition to Elasticsearch.

Data Usage : Unified console for visualization, alerts, and downstream integration.

Deploying the Agent on TKE

Because services run on Tencent Cloud TKE, the agent is stored on an NFS volume. The deployment steps are:

Step 1 : Modify the Docker start command to include the agent path.

$ java -javaagent:/nfs_data/XXX/agent/skywalking-agent.jar -Dspring.profiles.active=dev -jar target/youshu-app.jar

Step 2 : Upload the open‑source agent package to the NFS directory (e.g., /nfs_data/XXX/agent/).

Step 3 : Edit the configuration file config/agent.conf with the cloud‑provided collector address, token, and service name.

# Tencent Cloud TSW default collector address
collectorXXXXXX
# Unique token assigned to the cloud account
agent.XXXX@
# Custom service name
agent.service_name=XXX-api

Step 4 : Restart the service and verify topology and trace data appear in the UI.

APM Component Comparison

A comparison table (omitted here) shows why SkyWalking was chosen over other APM solutions, highlighting its open‑source nature, extensibility, and seamless integration with TSW.

Agent Probe Performance Impact

Official benchmarks indicate that enabling the Java Agent adds roughly a 10% CPU overhead for a typical web application, which is acceptable for most production environments.

Conclusion

The article demonstrates how Tencent Smart Retail solved three major challenges—service call monitoring, traceability, and performance diagnosis—by adopting Apache SkyWalking and the cloud‑native TSW platform, providing a practical reference for building observability in distributed microservice systems.

References

Detailed Java Agent guide: https://zhuanlan.zhihu.com/p/135872794

Tencent Cloud Microservice Observability Platform overview: https://cloud.tencent.com/document/product/1311/50754

Agent performance benchmark: https://github.com/SkyAPMTest/Agent-Benchmarks/blob/master/README_zh.md

SkyWalking project site: http://skywalking.apache.org/