TapTap runs inference workloads across search, advertising, and recommendation—and needed to launch new services in minutes, not days. By migrating its estimation platform to Serverless App Engine (SAE), TapTap eliminated manual infrastructure management and cut deployment time to minutes.
Background
XD Inc., founded in 2003, is a global game developer and publisher. As of mid-2022, XD operated 38 free-to-play and paid games. In 2016, XD launched TapTap—a mobile game community and app store where players download free-to-play and paid games and connect with other players. As of June 2022, TapTap had over 50 million monthly active users worldwide.
Unlike traditional app stores that take a revenue share, TapTap charges zero commission. Its revenue model depends on native in-app advertising—ads that match non-commercial content in style and format. These ads appear across five placements: game recommendations on the home page, content recommendations on the discovery page, hint words on the search guidance page, real-time search suggestions as users type, and the landing page after a search completes.
Three AI-driven workloads power these placements: automated deep learning model updates and deployments for search, advertising, and recommendation; a centralized model experiment recording platform for algorithm engineers; and natural language processing (NLP) analysis for new in-app content.
Challenges
TapTap originally deployed most backend services on Elastic Compute Service (ECS) instances managed with Rundeck. This constrained deployment efficiency and operational control. The following table summarizes TapTap's infrastructure challenges and how SAE addressed each one.
| Challenge | How SAE helped |
|---|---|
| Deploying and maintaining microservices infrastructure required dedicated headcount and slowed iteration | SAE delivers full microservices capabilities out of the box: automatic image building, built-in canary release, traffic shaping, environment isolation, application monitoring and diagnostics, and permission management. |
| Running Spring Cloud and Dubbo services at scale without a dedicated platform team | SAE supports Spring Cloud and Dubbo. Its fully managed microservices eliminate complex infrastructure setup and maintenance. |
| Managing ZooKeeper, Eureka, Consul, and SkyWalking added operational overhead on top of the core product | SAE manages the underlying stack—IaaS, Kubernetes, microservices components, and APM components—so teams do not need to deploy or maintain ZooKeeper, Eureka, Consul, or SkyWalking. SAE guarantees commercial-grade stability. |
| Launching Go services alongside Java required custom tooling and registration workarounds | SAE supports Go, Java, PHP, Python, and Node.js. Deploy Go applications as container images. Multi-language service registration and discovery runs on Kubernetes Services, with inter-service calls using Kubernetes Service names. |
Solution
The following figure shows TapTap's solution architecture. 
TapTap built a mid-tier estimation service on SAE that unifies model inference, feature development, and sample backtracking. All business lines—search, recommendation, and advertising—access this service at low cost. Concrete examples include click-through rate (CTR) estimation for search suggestions and CTR estimation for games in the International Edition.
The estimation service uses the Dubbo Go framework, which provides service registration and discovery, direct IP-based calls, and graceful startup and shutdown. SAE abstracts resource management, environment management, and O&M components entirely—enabling TapTap to launch independent estimation services for new scenarios and business lines in minutes.
TapTap also connected SAE's alerting platform, Event Center, and Simple Log Service. DingTalk alerts notify teams in real time when out-of-memory (OOM) errors, restarts, or error logs occur in live services.
Results
Minute-level releases. SAE supports canary release and rolling release. Integrating the OpenAPI with Git enables fast, automated deployments—new versions ship in minutes, regardless of business line.
Developers own the full delivery lifecycle. From development and release to ongoing O&M, developers handle the entire process end to end. Infrastructure concerns move off their plate entirely.
Scaling in seconds. SAE auto scaling responds to CPU, memory, queries per second (QPS), response time (RT), and scheduled triggers. Fine-tuned scaling policies reduce server costs as business grows.
Multi-language microservices at low cost. SAE provides runtimes for Java, PHP, Python, and Go. Kubernetes Service-based multi-language service discovery makes Go microservices straightforward to operate.