Catwalk的演变：Grab的模型服务平台

As Southeast Asia’s leading super app, Grab serves millions of users across multiple countries every day. Our services range from ride-hailing and food delivery to digital payments and much more. The backbone of our operations? Machine Learning (ML) models. They power our real-time decision-making capabilities, enabling us to provide a seamless and personalised experience to our users. Whether it’s determining the most efficient route for a ride, suggesting a food outlet based on a user’s preference, or detecting fraudulent transactions, ML models are at the forefront.

作为东南亚领先的超级应用，Grab每天为数百万用户提供服务。我们的服务范围从打车和外卖到数字支付等等。我们运营的支柱是什么？机器学习（ML）模型。它们为我们的实时决策能力提供动力，使我们能够为用户提供无缝和个性化的体验。无论是确定乘车的最高效路线，根据用户的偏好推荐食品店，还是检测欺诈交易，ML模型都处于前沿。

However, serving these ML models at Grab’s scale is no small feat. It requires a robust, efficient, and scalable model serving platform, which is where our ML model serving platform, Catwalk, comes in.

然而，在Grab的规模下提供这些机器学习模型并非易事。这需要一个强大、高效、可扩展的模型服务平台，这就是我们的机器学习模型服务平台Catwalk的用武之地。

Catwalk has evolved over time, adapting to the growing needs of our business and the ever-changing tech landscape. It has been a journey of continuous learning and improvement, with each step bringing new challenges and opportunities.

Catwalk随着时间的推移不断发展，适应了我们业务的不断增长和不断变化的技术环境。这是一段持续学习和改进的旅程，每一步都带来了新的挑战和机遇。

Evolution of the platform

平台的演进

Before Catwalk’s debut as our dedicated model serving platform, data scientists across the company employed various ad-hoc approaches to serve ML models. These included:

在Catwalk作为我们专用的模型服务平台之前，公司中的数据科学家采用了各种临时方法来提供ML模型的服务。其中包括：

Shipping models online using custom solutions.
使用自定义解决方案在线发布模型。
Relying on backend engineering teams to deploy and manage trained ML models.
依赖后端工程团队来部署和管理训练好的机器学习模型。
Embedding ML logic within Go backend services.
将ML逻辑嵌入Go后端服务中。

These methods, however, led to several challenges, undercovering the need for a unified, company-wide platform for serving machine learning models:

然而，这些方法带来了一些挑战，揭示了需要一个统一的、全公司范围内的机器学习模型服务平台的需求：

Operational overhead: Data scientists oft...