How Keras 3 Is Helping Expedia Modernize Its Ranking System

The Engineering Upgrade Behind a Tweet

Recently, a piece of information about Keras 3 helping Expedia modernize its ranking technology stack made the rounds in tech circles. While the original source was just a brief shared link, it points to a topic that's highly representative of today's machine learning engineering landscape: how large enterprises can leverage next-generation deep learning frameworks to rebuild the recommendation and ranking systems that power their core business.

For an online travel platform like Expedia, the ranking system is essentially its lifeblood — the order in which users see hotels and flights directly impacts conversion rates and revenue. Any modernization effort on the ranking stack is therefore an engineering decision with far-reaching consequences.

Why Keras 3

Keras 3 is a major rewrite of the Keras framework, with its standout feature being multi-backend support: the same model code can freely switch between TensorFlow, PyTorch, and JAX. This capability is particularly valuable for large enterprises, as it means teams are no longer locked into a single framework and can flexibly choose the underlying engine based on training efficiency, deployment environment, or team preferences.

For companies like Expedia that carry historical technical debt, framework migrations are often prohibitively expensive. The compatibility and unified API that Keras 3 provides can theoretically reduce the friction of migrating from legacy systems significantly, allowing engineering teams to progressively replace underlying implementations while preserving business logic.

Typical Pain Points in Ranking System Modernization

Enterprise-grade ranking systems typically accumulate several classes of problems over long iteration cycles: fragmented training and inference codebases, outdated framework versions that are difficult to upgrade, slow experimentation cycles, and difficulty in rapidly deploying new model architectures. These issues are further amplified in high-frequency transaction scenarios like travel and e-commerce.

When modernizing a ranking stack with a modern framework, the core objectives are usually to improve experimentation velocity and model maintainability — enabling data scientists to try new ranking algorithms faster and ship them safely to production.

Keras 3 was officially released in late 2023 (following a lengthy preview period) and represents a complete rewrite of the original Keras rather than an incremental upgrade. The long-standing coexistence of tf.keras and the standalone Keras version had caused significant community fragmentation. Keras 3 addresses this by introducing a unified backend abstraction layer — developers simply call unified APIs like keras.layers and keras.Model, while the construction and execution of the underlying computation graph is delegated to the chosen backend. Switching backends requires only setting an environment variable like KERAS_BACKEND=torch or KERAS_BACKEND=jax, with no changes needed to the model code itself. In practice, this means: during model development, you can use JAX's jit compilation for maximum training speed; during deployment, you can switch to TensorFlow Serving's mature inference infrastructure — all while keeping the same model definition code throughout.

Enterprise-grade recommendation and ranking systems typically adopt the Learning to Rank (LTR) paradigm, with common approaches divided into Pointwise (independently scoring each candidate), Pairwise (comparing relative order between pairs), and Listwise (directly optimizing the quality of the entire ranked list). Travel platforms like Expedia face the particular challenge of heterogeneous ranking signals: hotel rankings must simultaneously incorporate user historical behavior, price elasticity, inventory status, seasonal patterns, and other highly varied features. Traditional GBDT (Gradient Boosted Decision Tree)-based LTR approaches have inherent limitations in feature crossing and sequential behavior modeling — a key driver behind the push toward deep learning ranking architectures.

What Framework Choice Means for Engineering Teams

Keras 3's value lies not only in its technical features, but also in the unified language it provides for team collaboration. When research teams prefer JAX for its high performance while production teams rely on TensorFlow's mature deployment ecosystem, code that runs across backends can significantly ease collaboration friction.

This also reflects that competition among deep learning frameworks has entered a new phase: raw performance is no longer enough. Developer experience, migration costs, and ecosystem compatibility have become the critical considerations for enterprise framework selection. Expedia's case (as described in the original post) is a telling example of this trend.

JAX was developed by Google Research, and its core strengths lie in its functional programming style and the composability of jit (just-in-time compilation), vmap (vectorized mapping), and grad (automatic differentiation). In large-batch matrix operations and distributed training scenarios, it often achieves higher hardware utilization than PyTorch and TensorFlow, making it popular in academic research and performance-critical training settings. TensorFlow holds a clear advantage in the maturity of its production deployment ecosystem (TF Serving, TFLite, TF.js). PyTorch, with its dynamic graphs and rich community ecosystem, has become the mainstream choice for industrial ML development. Keras 3's multi-backend strategy is fundamentally about building bridges between these three ecosystems — each with distinct strengths — allowing enterprises to draw from each as needed, rather than making an all-or-nothing bet on a single framework.

Reference Value for the Industry

It's worth noting that the original source material for this article is limited — just a link to an Expedia engineering share. Specific technical details, performance improvement metrics, and migration practices should be sought in the full original source.

That said, this type of enterprise case study still offers meaningful lessons for ML engineering teams broadly: in a period of rapid framework evolution, choosing a framework with multi-backend capability and migration-friendly design may be the prudent strategy for managing future uncertainty. For teams currently evaluating a technology stack upgrade, Expedia's path is well worth continued attention.