Apache Beam’s latest release, version 0.6.0, introduces a new SDK – this time, for the Python programming language. The Python SDK joins the Java SDK as the second implementation of the Beam programming model.
The Python SDK incorporates all of the main concepts of the Beam model, including ParDo, GroupByKey, Windowing, and others. It features extensible IO APIs for writing bounded sources and sinks, and provides built-in implementation for reading and writing Text, Avro, and TensorFlow record files, as well as connectors to Google BigQuery and Google Cloud Datastore.
There are two runners capable of executing pipelines written with the Python SDK today: Direct Runner and Dataflow Runner, both of which are currently limited to batch execution only. Upcoming features will shortly bring the benefits of the Python SDK to additional runners.
Try the Apache Beam Python SDK
Let’s take a look at a quick example together. First, install the
apache-beam package from PyPI and start your Python interpreter.
$ pip install apache-beam $ python
We will harness the power of Apache Beam to estimate Pi in honor of the recently passed Pi Day.
import random import apache_beam as beam def run_trials(count): """Throw darts into unit square and count how many fall into unit circle.""" inside = 0 for _ in xrange(count): x, y = random.uniform(0, 1), random.uniform(0, 1) inside += 1 if x*x + y*y <= 1.0 else 0 return count, inside def combine_results(results): """Given all the trial results, estimate pi.""" total, inside = sum(r for r in results), sum(r for r in results) return total, inside, 4 * float(inside) / total if total > 0 else 0 p = beam.Pipeline() (p | beam.Create( * 10) # Create 10 experiments with 500 samples each. | beam.Map(run_trials) # Run experiments in parallel. | beam.CombineGlobally(combine_results) # Combine the results. | beam.io.WriteToText('./pi_estimate.txt')) # Write PI estimate to a file. p.run()
This example estimates Pi by throwing random darts into the unit square and keeping track of the fraction of those darts that fell into the unit circle (see the full example for details). If you are curious, you can check the result of our estimation by looking at the output file.
$ cat pi_estimate.txt*
The first thing on the Python SDK’s roadmap is to address two of its limitations. First, the existing runners are currently limited to bounded PCollections, and we are looking forward to extending the SDK to support unbounded PCollections (“streaming”). Additionally, we are working on extending support to more Apache Beam runners, and the upcoming Fn API will do the heavy lifting.
Both of these improvements will enable the Python SDK to fulfill the mission of Apache Beam: a unified programming model for batch and streaming data processing that can run on any execution engine.
Please consider joining us, whether as a user or a contributor, as we work towards our first release with API stability. If you’d like to try out Apache Beam today, check out the latest 0.6.0 release. We welcome contributions and participation from anyone through our mailing lists, issue tracker, pull requests, and events.
文章来源: Python SDK released in Apache Beam 0.6.0