Since 2018, Amber Mobile has relied on a range of services provided by Google Cloud to build and support its technological infrastructure. In 2021, Amber Mobile began utilizing fully managed Google Cloud services such as BigQuery, Cloud Storage, and Dataproc to implement the modernization of its data lake and data warehouse.

Cloud Storage provides Amber Mobile with a cost-effective and high-performance data storage solution, supporting the development and deployment of data pipelines with a robust data lake foundation. As a compute engine, BigQuery can directly access and process data stored in Cloud Storage. In the process of creating a data warehouse, the calculation results of Dataproc can be seamlessly connected. This convenient interoperability and simplified integration methods reduce the management difficulty of data lakes and warehouses for data teams. By migrating from a traditional, self-built infrastructure using tightly-coupled open-source products to Google Cloud's managed services, Amber Mobile's data team has achieved module decoupling. The platform's auto-scaling capabilities have significantly improved query and analysis performance.

For workflow orchestration and scheduling, Amber Mobile uses Cloud Composer to flexibly implement pipeline orchestration. Since this service retains many features of the open-source Apache Airflow, Amber Mobile’s data team can write code in familiar languages and integrate it end-to-end with BigQuery, Dataproc, and Cloud Storage. This enables precise control over data flow and result generation.

In the modernization of its data processing, Dataproc empowers Amber Mobile’s data team to support over 30 open-source tools and frameworks, including Apache Hadoop, Apache Spark, and Apache Flink, through its fully managed, serverless service. The option to incorporate Apache Hudi as the data lake foundation further ensures comprehensive open-source compatibility. This modernization significantly enhances data processing efficiency, enabling the data team to handle complex data structures and scheduling logic with ease. Leveraging Dataproc and BigQuery’s ability to directly access data stored in Cloud Storage, Amber Mobile can efficiently process and store diverse data sources, server-side data, and metrics with complex interconnections across applications in the current data warehouse, then visualize data in Looker Studio, and uncover value hidden within the data through rapid analysis, providing valuable guidance for product optimization and iteration.

Amber Mobile has been using Google Cloud’s innovations in AI and multimodal models, and leveraging AI to realize product features and enhance user experiences.

When building AI models, Amber Mobile leverages Google Cloud’s Jupyter Notebook-compatible environment provided in services such as BigQuery, Dataproc, and Vertex AI to seamlessly integrate data and model resources with Notebook. This enables internal team sharing and collaboration, facilitating more efficient model building and testing. With Google Cloud’s intuitive tools and straightforward configurations, Amber Mobile can transition code from the development environment to production with minimal modifications. This streamlined process significantly reduces barriers between development and operations while markedly easing the workload.

In terms of continuous delivery for MLOps, Amber Mobile leverages multiple efficient components from Google Cloud to accurately predict user value and behavior, enabling seamless integration of machine learning workflows from development to operations. The process starts with large-scale data storage and analysis in cloud-based data warehouses, ensuring high efficiency and scalability in data processing. Amber Mobile then uses the AI platform to train machine learning models, which supports frameworks like TensorFlow and provides automated hyperparameter tuning alongside parameter scales ranging from millions to tens of millions, significantly enhancing model prediction accuracy. Once trained, the models are deployed directly to BigQuery ML for offline batch processing tasks. By precisely predicting user behavior, Amber Mobile can design game operations that better align with user habits and preferences, increasing user satisfaction and loyalty while driving business growth.

In Computer Vision, Amber Mobile not only leverages cloud-based computing resources for customized model training but also adopts Document AI to automate the extraction of information from both unstructured and structured documents. This includes extracting content relevant to optimizing game experience from files such as screenshots, scanned documents, photos, and PDFs, and combining the extracted data with contextual information to gain deeper insights. Previously, processing large volumes of low-quality documents required significant effort from the Amber Mobile team. Now, with Document AI’s built-in foundational models, the team can effectively identify standard content with an impressive accuracy of up to 99.9%. These models can also be fine-tuned and automatically tagged based on specific business needs, allowing the team to handle high volumes of tasks with ease and efficiency.

For years, Amber Mobile has relied on Google Kubernetes Engine (GKE) for its flexible resource management to support business deployments. To facilitate large-scale model training, Amber Mobile has also chosen GKE for distributed AI training and application deployment. Notably, in reinforcement training scenarios, Amber Mobile leverages GKE’s support for Ray, using Ray on GKE to realize large-scale sampling in reinforcement learning. This facilitates rapid deployment and horizontal scaling of training tasks without the need for complex code modifications or configuration adjustments, ensuring adequate samples for AI training.