Load Data Into Google BigQuery and AutoML
In this blog, we will review ETL data pipeline in StreamSets Transformer, a Spark ETL engine, to ingest real-world data from Fire Department of New York (FDNY) stored in Google Cloud Storage (GCS), transform it, and load data in Google BigQuery curated.
Once the transformed data is made available in Google BigQuery, it will be used in AutoML to train a machine learning model to predict the average incident response time for the FDNY.
What is Dataproc?
Dataproc is a low-cost, Google Cloud Platform integrated, easy to use managed Spark and Hadoop service that can be leveraged for batch processing, streaming, and machine learning use cases.
What is Google BigQuery?
BigQuery is an enterprise grade data warehouse that enables high-performance SQL queries using the processing power of Google’s infrastructure.
The dataset is made available through the NYC Open Data website. The 2009-2018 historical dataset contains average response times of the FDNY. The data is partitioned by incident type (False Alarm, Medical Emergency, and so on), borough, and the number of incidents during a particular month.
Here’s what the sample FDNY data looks like:
YEARMONTH,INCIDENTCLASSIFICATION,INCIDENTBOROUGH,INCIDENTCOUNT,AVERAGERESPONSETIME 2009/07,All Fire/Emergency Incidents,Citywide,40850,04:27 2009/07,False Alarm,Citywide,2655,04:07 2009/07,Medical Emergencies,Manhattan,4895,04:17 2009/07,Medical False Alarm,Citywide,408,04:13 2009/07,NonMedical Emergencies,Manhattan,4446,04:51 2009/07,NonStructural Fires,Citywide,1495,04:19 2009/08,All Fire/Emergency Incidents,Citywide,41723,04:30
Data Pipeline Overview
Data Source And Dataset
- Data in CSV format is loaded from GCS using Google Cloud Storage (GCS) origin. To load data from GCS, all you need to provide is the path to the bucket, data format, and file name pattern.
- Raw data is transformed using Filter, Field Remover, and Spark SQL Expression processors in a format that is suitable for machine learning. (See details below.)
- Transformed data is stored in a Google BigQuery table. Note: if the table doesn’t already exist, it will be created automatically by StreamSets Transformer.
- In this example, the data pipeline is designed to run on an existing or ephemeral Google Dataproc cluster. Note: Other supported cluster types in StreamSets Transformer include Databricks, Amazon EMR, Azure for HDInsight, Hadoop YARN, and SQL Server 2019 Big Data Cluster.
Load Data Into Google BigQuery and AutoML | Data Pipeline Preview
Before running the Spark ETL pipeline in StreamSets Transformer, you can preview the pipeline against the configured Dataproc cluster to examine the data structure, data types, and verify the transformations at every stage. This is also a great way to debug data pipelines. For more information on pipeline preview, refer to the documentation.
Using a Filter processor we will filter out incidents where INCIDENTCLASSIFICATION == “All Fire/Emergency Incidents“ or INCIDENTBOROUGH == “Citywide”.
Remove Future Information
Because this is a historical dataset and we’re using it to train a machine learning model, we need to remove information that would not be known at the beginning of the month. In this case, that is INCIDENTCOUNT. To remove this field from every record, we’ll use a Field Remover processor.
Labels or target variables in machine learning models are of numeric data type. In this case, the field value of AVERAGERESPONSETIME is transformed in the following steps:
- Remove “:” using Spark SQL expression — replace(AVERAGERESPONSETIME,”:”,””)
- Convert from time to seconds and from string datatype to integer using Spark SQL expression — round((AVERAGERESPONSETIME / 100) * 60 + (AVERAGERESPONSETIME % 100))
StreamSets enables data engineers to build end-to-end smart data pipelines. Spend your time building, enabling and innovating instead of maintaining, rewriting and fixing.
Data Pipeline Execution
Running the StreamSets Transformer data pipeline displays various metrics in real-time. For example, batch processing time taken by each stage as shown below. This is a great way to start looking into fine tuning the processing and transformations.
Load Data into Google BigQuery
Once the pipeline runs successfully, the Google BigQuery table is auto-created, if it doesn’t already exists, and the transformed data is inserted into the table. This dataset is then readily available for querying as shown below.
The transformed data stored can then be imported directly from the BigQuery table for training a machine learning model in AutoML.
Using AutoML you can build on Google’s machine learning capabilities and create custom machine learning models.
Select Target Column
Train Machine Learning Model
That’s it! We went from loading raw, real-world data into Google BigQuery to creating a machine learning model in AutoML without any coding or scripting!
Build Your Spark ETL and ML Data Pipelines
It goes without saying that training models, evaluating them, model versioning, and serving different versions of the model are non-trivial undertakings and that is not the focus of this post. That said, however, StreamSets Transformer makes it really easy to load data into Google BigQuery and AutoML.
Checkout these helpful resources and get started quickly with running your Spark ETL data pipelines. Here are some other technical blogs related to Machine Learning that you might be interested in reading:
- Model Experiments, Tracking and Registration using MLflow on Databricks
- Streaming Analysis Using Spark ML in StreamSets DataOps Platform
- StreamSets Transformer: Natural Language Processing in PySpark
- StreamSets Transformer Extensibility: Spark MLeap Bundles to S3
Learn more about StreamSets For Google Cloud Platform.