Data Science on the Google Cloud Platform, 2nd Edition

Chapter 9. Machine Learning with TensorFlow in Vertex AI

In Chapter 7, we built a machine learning model in Spark but ran into problems when trying to scale it out and make it operational. We were able to address the scalability challenge by using BigQuery ML in Chapter 8, but the operationalization challenges still remain. In addition, although BigQuery ML was scalable, we were not able to build the most expressive ML model possible. Briefly, there are four challenges that we identified:

• One-hot encoding of categorical columns caused an explosion in the size of the dataset because of the increased size of the columns. BigQuery ML was able to handle this, but Spark wasn’t.

• Embeddings would have involved special bookkeeping in Spark, and this was not an option in BigQuery ML.

• Putting the model into production requires the machine learning library to be portable to environments beyond the Hadoop cluster or BigQuery data warehouse on which the model is trained.

• Preventing training–serving skew when using a time-windowed aggregate feature requires being able to use the same data preparation code for both historical data (which is batch) and real-time data (which is streaming).

We will solve the fourth problem, of time-windowed aggregates, in Chapter 11 by using Apache Beam and its ability to employ the same code for both batch and stream.

The solution to the first three problems requires a portable machine learning library that is (1) powerful enough to carry out training using ...