TFX

Abstract

Creating and maintaining a platform for reliably producing and deploying machine learning models requires careful orchestration of many components-a learner for generating models based on training data, modules for analyzing and validating both data as well as models, and finally infrastructure for serving models in production. This becomes particularly challenging when data changes over time and fresh models need to be produced continuously. Unfortunately, such orchestration is often done ad hoc using glue code and custom scripts developed by individual teams for specific use cases, leading to duplicated effort and fragile systems with high technical debt.We present TensorFlow Extended (TFX), a TensorFlowbased general-purpose machine learning platform implemented at Google. By integrating the aforementioned components into one platform, we were able to standardize the components, simplify the platform configuration, and reduce the time to production from the order of months to weeks, while providing platform stability that minimizes disruptions.We present the case study of one deployment of TFX in the Google Play app store, where the machine learning models are refreshed continuously as new data arrive. Deploying TFX led to reduced custom code, faster experiment cycles, and a 2% increase in app installs resulting from improved data and model analysis.

MLIoT

“…Several recent efforts aim to simplify ML development through a general-purpose machine learning system with both training and serving of models [2,5,6,13,15,26,34,35,42,68,69]. Some of these systems that share similar goals of MLIoT are the end-toend "ML Platforms" that run at commercial settings.…”

“…FBLearner Flow allows reusable ML workflow that allows ML models to be modified and reused in different products. On the other hand, Google's TFX [6], provides Tensorflow-based [24] toolkits for data preparation, periodic model evaluation to improve performance and reliability and extends TensorFlow Serving [51] to serve the models with TensorFlowbased learners. Such systems generally run on the cloud incurring a higher cost for better workload environments and restrict users to a specific set of algorithms or libraries, so users are on their own when they step outside these boundaries.…”

“…There is widespread interest in Internet-of-Things (IoT) enabled environments with the expectation that it can revolutionize healthcare, smart homes/buildings, and manufacturing. This vision is enabled by IoT sensors with rich sensing capabilities [39,49], cloud platforms like AWS and Azure, and rapid advances in ML software stacks [4][5][6]. In the context of smart homes and buildings, in particular, the goal is to sense the rich context and activities of the occupants and to detect the events/states of the appliances and equipment in the space [32,38,39].However, supporting the diversity of IoT use cases, differences in user and application requirements, as well as the dynamic environments that these devices are deployed in, requires machine learning platforms that can be customized and adapted to the IoT domain.…”

Towards ML Engineering with TensorFlow Extended (TFX)

“…The popularity and impact of TensorFlow [9] within and outside of Alphabet, the popularity and impact of TFX within Alphabet, and the reality that equivalents of ML engineering will be needed by organizations and individuals everywhere in the world, felt like something we could not ignore. That led us to publicly describe the design and initial deployment of TFX within Google [10] and to, step by step, make more of our learnings and our technology publicly available (including open source), while we continue building more of each. We were able to accomplish this in part because, like Sibyl, TFX built upon robust infrastructural dependencies.…”

“…This realization affected the implementation and evolution of Sibyl; it was entrenched in TFX by the time we publicly wrote about it [10] and was ultimately generalized and formalized in ML Metadata [19], now powering TFX.…”