IT Operations Automation · Marketing Personalization
TapTap's ad-driven recommendation, search, and content workflows depended on machine-learning models and NLP processing, but the older ECS-plus-Rundeck environment made deployments and service management too slow and expensive. The team needed a more elastic platform for automated model updates, experiment tracking, content parsing, and online estimation services while continuing to support Go-based services and fast release cycles for a globally used game community.
Reported impact
TapTap gave its algorithm engineers a faster, lower-ops platform for model deployment, NLP enrichment, and online estimation. The serverless design let the company decouple triggering logic from business code, reduce infrastructure management, and support faster releases and elastic scaling for recommendation and advertising services that affect user experience and monetization.
XD moved TapTap's algorithm platform toward an all-on-serverless design using Function Compute (FC) and Serverless Application Engine (SAE). Model artifacts exported into OSS buckets trigger FC deployment workflows, so TensorFlow, PyTorch, and similar models can be published without separate engineering handoffs. An internal experiment-management service also runs on FC with HTTP triggers, recording training parameters and log addresses so model experiments stay traceable. New user posts are pushed through Kafka to an NLP parsing service, where event-driven execution invokes analysis only when content arrives and stores the results for later search use. The team also schedules FC jobs to pull unstructured billing data and aggregate MaxCompute and EAS consumption by task and model. On the serving side, SAE hosts shared estimation services used by search, recommendation, and advertising workloads, giving TapTap faster releases, microservice support, and lower operations overhead.
Model artifacts are written to OSS and trigger FC-based deployment jobs, experiment submissions arrive through HTTP-triggered FC services that log parameters and log locations, and new user posts are pushed through Kafka into event-driven NLP analysis for search indexing. Processed signals and online prediction services are then served through SAE, while scheduled FC tasks aggregate MaxCompute and EAS consumption data for internal cost control.
New posts are pushed through Kafka to NLP processing for later search use, and SAE gave TapTap minute-level releases plus second-level elastic scaling. (original: 当 TapTap 站内的用户发了一个新的帖子，会通过 Kafka 推送到 NLP 分析服务商进行 NLP 的处理与解析，存下来用于之后的搜索……不停机发布 + 分钟级上线……秒级弹性缩扩容。)
TapTap said its serverless work was driven by algorithm scenarios such as deep-learning-model updates and deployment, model-experiment records, and NLP analysis for new in-platform content. (original: 搜广推都依赖的深度学习模型自动化更新/部署，以及组内算法工程师都需要依赖的模型实验记录平台，还有站内新内容的一些 NLP 分析处理等。)
The team chose a Serverless architecture built from SAE and Function Compute because it could cover most business scenarios with lower operations cost than ECS plus self-built Kubernetes. (original: 心动技术团队选择了……Serverless 架构，使用 Serveless 应用引擎（SAE）和函数计算 FC……可以覆盖掉 TapTap 绝大多数的业务场景，实现所有应用服务 All On Serverless。)
When model files are exported to an OSS bucket, TapTap automatically triggers deployment and update services so algorithm engineers can publish models without depending on extra engineering staff. (original: 算法工程师在训练完自己的模型……只需要导出到指定的 OSS 存储空间 Bucket，就会触发模型的更新与部署服务。这样算法工程师在不依赖其他工程人力的情况下也能执行模型的部署、更新以及后续的弹性缩扩容。)
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
Related
Alibaba Cloud Sample Oct 12, 2026 ECLICKTECH (易点天下) ECLICKTECH needed to industrialize AI-driven advertising operations, but media-delivery workflows were still long, manual, and heavily dependent on senior operators for anomaly diagnosis and creative adaptation. As more… AI covers about 50% of core execution work AdClaw · AgentCore · AgentLoop 
Google Cloud Sample Archived · Sep 25, 2026 TapNation However, the company quickly realized that the true revolution lay elsewhere. "For us, AI isn't just another tool stacked on top of existing ones. It is the backbone of a new operating model, capable of supporting all… 60% reduction in time spent drafting technical specifications Gemini · Gemini Enterprise · Gemini Enterprise Agent Platform Google Cloud Sample Archived · Sep 9, 2026 The Points Guy As teams shift from generative AI experimentation to production, scaling video creative assets has emerged as a major operational bottleneck. While ad platforms allow for strategic audience targeting, the production of… 71,000+ clicks and 67,000+ conversions driven by GenAI video assets Gemini · Veo 