Telemetry/Custom analysis with spark

Introduction

Spark is a data processing engine designed to be fast and easy to use. We have setup Jupyter workbooks that use Spark to analyze our Telemetry data. Jupyter workbooks can be easily shared and updated among colleagues to enable richer analysis than SQL alone.

The Spark clusters can be spun up on analysis.telemetry.mozilla.org, which is abbreviated as atmo. The Spark Python API is called pyspark.

Setting Up a Spark Cluster On atmo

1. Go to analysis.telemetry.mozilla.org
2. Click “Launch an ad-hoc Spark cluster”.
3. Enter some details:
   1. The “Cluster Name” field should be a short descriptive name, like “chromehangs analysis”.
   2. Set the number of workers for the cluster. Please keep in mind to use resources sparingly; use a single worker to write and debug your job.
   3. Upload your SSH public key.
4. Click “Submit”.
5. A cluster will be launched on AWS preconfigured with Spark, IPython and some handy data analysis libraries like pandas and matplotlib.

Once the cluster is ready, you can tunnel IPython through SSH by following the instructions on the dashboard, and running the ssh shell command. For example:

ssh -i ~/.ssh/id_rsa -L 8888:localhost:8888 hadoop@ec2-54-70-129-221.us-west-2.compute.amazonaws.com

Finally, you can launch IPython in Firefox by visiting http://localhost:8888.

The Python Jupyter Notebook

When you access http://localhost:8888, two example Jupyter notebooks are available to peruse. To create a new Jupyter notebook, select new -> Python 2.

Starting out, we recommend looking through the Telemetry Hello World notebook. It gives a nice overview of Jupyter and analyzing telemetry data using pyspark and the RDD API.

Using Jupyter

Jupyter Notebooks contain a series of cells. Each cell contains code or markdown. To switch between the two, use the dropdown at the top. To run a cell, use shift-enter; this either compiles the markdown or runs the code. To create new cell, select Insert -> Insert Cell Below.

A cell can output text or plots. To output plots inlined with the cell, run the following command, usually below your import statements:

 %pylab inline

The notebook is setup to work with Spark. See the "Using Spark" section for more information.

Schedule a periodic job

Scheduled Spark jobs allow a Jupyter notebook to be updated consistently, making a nice and easy-to-use dashboard.

To schedule a Spark job:

1. Visit the analysis provisioning dashboard at telemetry-dash.mozilla.org and sign in using Persona with an @mozilla.com email address.
2. Click “Schedule a Spark Job”.
3. Enter some details:
   1. The “Job Name” field should be a short descriptive name, like “chromehangs analysis”.
   2. Upload your IPython notebook containing the analysis.
   3. Set the number of workers of the cluster in the “Cluster Size” field.
   4. Set a schedule frequency using the remaining fields.

Now, the notebook will be updated automatically and the results can be easily shared. Furthermore, all files stored in the notebook's local working directory at the end of the job will be automatically uploaded to S3, which comes in handy for simple ETL workloads for example.

For reference, see Simple Dashboard with Scheduled Spark Jobs and Plotly.

Sharing a Notebook

Jupyter notebooks can be shared in a few different ways.

Sharing a Static Notebook

An easy way to share is using a gist on github.

1. Download file as .ipynb
2. Upload to a gist on gist.github.com
3. Enter the gist URL at Jupyter nbviewer
4. Share with your colleagues!

Sharing a Scheduled Notebook

Setup your scheduled notebook. After it's run, do the following:

1. Go to the 'Schedule a Spark job' tab in atmo
2. Get the URL for the notebook (under 'Currently Scheduled Jobs')
3. Paste that URL into Jupyter nbviewer

Using Spark

Spark is a general-purpose cluster computing system - it allows users to run general execution graphs. APIs are available in Python, Scala, and Java. The Jupyter notebook utilizes the Python API. In a nutshell, it provides a way to run functional code (e.g. map, reduce, etc.) on large, distributed data.

Check out Spark Best Practices for tips on using Spark to it's full capabilities.

SparkContext (sc)

Access to the Spark API is provided through SparkContext. In the Jupyter notebook, this is the `sc` object. For example, to create a distributed RDD of monotonically increasing numbers 1-1000:

   numbers = range(1000)
   #no need to initialize sc in the Jupyter notebook
   numsRdd = sc.parallelize(numbers)
   nums.take(10) #no guaranteed order

Spark RDD

The Resilient Distributed Dataset (RDD) is Spark's basic data structure. The operations that are performed on these structures are distributed to the cluster. Only certain actions (such as collect() or take(N)) pull an RDD in locally.

RDD's are nice because there is no imposed schema - whatever they contain, they distribute around the cluster. Additionally, RDD's can be cached in memory, which can greatly improve performance of some algorithms that need access to data over and over again.

Additionally, RDD operations are all part of a directed, acyclic graph. This gives increased redundancy, since Spark is always able to recreate an RDD from the base data (by rerunning the graph), but also provides lazy evaluation. No computation is performed while an RDD is just being transformed (a la map), but when an action is taken (e.g. reduce, take) the entire computation graph is evaluated. Continuing from our previous example, the following gives some of the peaks of a sin wave:

   import numpy as np
   #no computation is performed on the following line!
   sin_values = numsRdd.map(lambda x : np.float(x) / 10).map(lambda x : (x, np.sin(x)))
   #now the entire computation graph is evaluated
   sin_values.takeOrdered(5, lambda x : -x[1])

For jumping into working with Spark RDD's, we recommend reading the Spark Programming Guide.

Spark SQL and Spark Dataframes/Datasets

Spark also supports traditional SQL, along with special data structures that require schemas. The Spark SQL API can be accessed with `sqlContext`. For example:

   longitudinal = sqlContext.sql('SELECT * FROM longitudinal')

creates a DataFrame that contains all the longitudinal data. A Spark DataFrame is essentially a distributed table, a la Pandas or R Dataframes. Under the covers they are an RDD of Row objects, and thus the entirety of the RDD API is available for DataFrames, as well as a DataFrame specific API. For example, a sql-like way to get the count of a specific OS:

   longitudinal.select("os").where("os = 'Darwin'").count()

To Transform the DataFrame object to an RDD, simply do:

  longitudinal_rdd = longitudinal.rdd

In general, however, the DataFrames are performance optimized, so it's worth the effort to learn the DataFrame API.

For more overview, see the SQL Programming Guide. See also the Longitudinal Tutorial, one of the available example notebooks when you start a cluster.

Available Data Sources for SparkSQL

For information about available queryable data sources (e.g. Longitudinal dataset), see Telemetry Datasets Documentation.

These datasets are optimized for fast access, and will far out-perform analysis on the raw telemetry ping data.

Accessing the Spark UI

Go to localhost:8888/spark after sshing into the spark cluster to see the Spark UI. It has information about job statuses and task completion, and may help you debug your job.

The MozTelemetry Library

We have provided a library that gives easy access to the raw telemetry ping data. For example usage, see the Telemetry Hello World example notebook. Detailed documentation for the library can be found at the Python MozTelemetry Documentation.

Using the Raw Ping Data

First off, import the moztelemetry library using the following:

   from moztelemetry import get_pings

If you need any other functions, just comma separate them with get_pings.

The ping data is an RDD of JSON elements. For example, using the following:

   pings = get_pings(sc, app="Firefox", channel="nightly", build_id=("20160901000000", "20160901999999"), fraction=0.01)

returns an RDD of 1/100th of Firefox Nightly JSON pings for all builds from September 1 2016. Now, because it's JSON, pings are easy to access. For example, to get the count of each OS type:

   os_names = pings.map(lambda x : (x['environment']['system']['os']['name'], 1))
   os_counts = os_names.reduceByKey(lambda x, y : x+y)
   os_counts.collect()

Alternatively, moztelemetry provides the `get_pings_properties` function, which will gather the data for you:

   subset = get_pings_properties(pings, ["environment/system/os/name"])
   subset.map(lambda x : (x["environment/system/os/name"], 1)).reduceByKey(lambda x, y : x+y).collect()

FAQ

Please add more FAQ as questions are answered by you or for you.

How can I load parquet datasets in a Jupyter notebook?

Use sqlContext.read.load, e.g.:

 dataset = sqlContext.read.load("s3://the_bucket/the_prefix/the_version", "parquet")

I got a REMOTE HOST IDENTIFICATION HAS CHANGED! error

AWS recycles hostnames, so removing the offending key from $HOME/.ssh/known_hosts will remove the warning. You can find the line to remove by finding the line in the output that says

 Offending key in /path/to/hosts/known_hosts:2

Where 2 is the line number of the key that can be deleted. Just remove that line, save the file, and try again.

Why is my notebook hanging?

There are a couple common causes for this:

1. Currently, our Spark notebooks can only run a single Python kernel at a time. If you open multiple notebooks on the same cluster and try to run both, the second notebook will hang. Be sure to close notebooks using "Close and Halt" under the "File" dropdown.

2. The connection from PySpark to the Spark driver might be lost. Unfortunately the best way to recover from this for the moment seems to be spinning up a new cluster.

3. Canceling execution of a notebook cell doesn't cancel any spark jobs that might be running in the background. If your spark commands seem to be hanging, try running `sc.cancelAllJobs()`.