# Complex Aggregations in PySpark

I’ve touched on this in past posts, but wanted to write a post specifically describing the power of what I call complex aggregations in PySpark.

The idea is that you have have a data request which initially seems to require multiple different queries, but using ‘complex aggregations’ you can create the requested data using a single query (and a single shuffle).

Let’s say you have a dataset like the following. You have one column (id) which is a unique key for each user, another column (group) which expresses the group that each user belongs to, and finally (value) which expresses the value of each customer. I apologize for the contrived example.

id group value
1 'a' 5.1
2 'b' 2.6
3 'b' 3.4
4 'c' 1.7

Let’s say someone wants the average value of group a, b, and c, AND the average value of users in group a OR b, the average value of users in group b OR c AND the value of users in group a OR c. Adds a wrinkle, right? The ‘or’ clauses prevent us from using a simple groupby, and we don’t want to have to write 4 different queries.

Using complex aggregations, we can access all these different conditions in a single query.

group_a_avg group_b_avg group_c_avg group_ab_avg group_ac_avg group_bc_avg
5.1 3.0 1.7 3.7 3.4 2.6

They key here is using when to filter different data in and out of different aggregations.

This approach can be quite concise when used with python list comprehensions. I’ll rewrite the query above, but using a list comprehension.

group_a_avg group_b_avg group_c_avg group_ab_avg group_ac_avg group_bc_avg
5.1 3.0 1.7 3.7 3.4 2.6

Voila! Hope you find this little trick helpful! Let me know if you have any questions or comments.

# Introducing Predeval

Predeval is software designed to help you identify changes in a model’s output.

For instance, you might be tasked with building a model to predict churn. When you deploy this model in production, you have to wait to learn which users churned in order to know how your model performed. While Predeval will not free you from this wait, it can provide initial signals as to whether the model is producing reasonable (i.e., expected) predictions. Unexpected predictions might reflect a poor performing model. They also might reflect a change in your input data. Either way, something has changed and you will want to investigate further.

Using predeval, you can detect changes in model output ASAP. You can then use python’s libraries to build a surrounding alerting system that will signal a need to investigate. This system should give you additional confidence that your model is performing reasonably. Here’s a post where I configure an alerting system using python, mailutils, and postfix (although the alerting system is not built around predeval).

Predeval operates by forming expectations about what your model’s outputs will look like. For example, you might give predeval the model’s output from a validation dataset. Predeval will then compare new outputs to the outputs produced by the validation dataset, and will report whether it detects a difference.

Predeval works with models producing both categorical and continuous outputs.

Here’s an example of predeval with a model producing categorical outputs. Predeval will (by default) check whether all expected output categories are present, and whether the output categories occur at their expected frequencies (using a Chi-square test of independence of variables in a contingency table).

Here’s an example of predeval with a model producing continuous outputs. Predeval will (by default) check whether the new output have a minimum lower than expected, a maximum greater than expected, a different mean, a different standard deviation, and whether the new output are distributed as expected (using a Kolmogorov-Smirnov test)

I’ve tried to come up with reasonable defaults for determining whether data are different, but you can also set these thresholds yourself. You can also choose what comparison tests to run (e.g., checking the minimum, maximum etc.).

You will likely need to save your predeval objects so that you can apply them to future data. Here’s an example of saving the objects.

Documentation about how to install predeval can be found here.

# Creating a Survival Function in PySpark

Traditionally, survival functions have been used in medical research to visualize the proportion of people who remain alive following a treatment. I often use them to understand the length of time between users creating and cancelling their subscription accounts.

Here, I describe how to create a survival function using PySpark. This is not a post about creating a Kaplan-Meier estimator or fitting mathematical functions to survival functions. Instead, I demonstrate how to acquire the data necessary for plotting a survival function.

I begin by creating a SparkContext.

Next, I load fake data into a Spark Dataframe. This is the data we will use in this example. Each row is a different user and the Dataframe has columns describing start and end dates for each user. start_date represents when a user created their account and end_date represents when a user canceled their account.

id start_date end_date
1 2018-11-01 2018-11-03
2 2018-01-01 2018-08-17
3 2017-12-31 2018-01-06
4 2018-11-15 2018-11-16
5 2018-04-02 2018-04-12

I use start_date and end_date to determine how many days each user was active following their start_date.

id start_date end_date days_till_cancel
1 2018-11-01 2018-11-03 2
2 2018-01-01 2018-08-17 228
3 2017-12-31 2018-01-06 6
4 2018-11-15 2018-11-16 1
5 2018-04-02 2018-04-12 10

I use a Python UDF to create a vector of the numbers 0 through 13 representing our period of interest. The start date of our period of interest is a user’s start_date. The end date of our period of interest is 13 days following a user’s start_date. I chose 13 days as the period of interest for no particular reason.

I use explode to expand the numbers in each vector (i.e., 0->13) into different rows. Each user now has a row for each day in the period of interest.

I describe one user’s data below.

id start_date end_date days_till_cancel day
1 2018-11-01 2018-11-03 2 0
1 2018-11-01 2018-11-03 2 1
1 2018-11-01 2018-11-03 2 2
1 2018-11-01 2018-11-03 2 3
1 2018-11-01 2018-11-03 2 4
1 2018-11-01 2018-11-03 2 5
1 2018-11-01 2018-11-03 2 6
1 2018-11-01 2018-11-03 2 7
1 2018-11-01 2018-11-03 2 8
1 2018-11-01 2018-11-03 2 9
1 2018-11-01 2018-11-03 2 10
1 2018-11-01 2018-11-03 2 11
1 2018-11-01 2018-11-03 2 12
1 2018-11-01 2018-11-03 2 13

We want the proportion of users who are active X days after start_date. I create a column active which represents whether users are active or not. I initially assign each user a 1 in each row (1 represents active). I then overwrite 1s with 0s after a user is no longer active. I determine that a user is no longer active by comparing the values in day and days_till_cancel. When day is greater than days_till_cancel, the user is no longer active.

I describe one user’s data below.

id start_date end_date days_till_cancel day active
1 2018-11-01 2018-11-03 2 0 1
1 2018-11-01 2018-11-03 2 1 1
1 2018-11-01 2018-11-03 2 2 0
1 2018-11-01 2018-11-03 2 3 0
1 2018-11-01 2018-11-03 2 4 0
1 2018-11-01 2018-11-03 2 5 0
1 2018-11-01 2018-11-03 2 6 0
1 2018-11-01 2018-11-03 2 7 0
1 2018-11-01 2018-11-03 2 8 0
1 2018-11-01 2018-11-03 2 9 0
1 2018-11-01 2018-11-03 2 10 0
1 2018-11-01 2018-11-03 2 11 0
1 2018-11-01 2018-11-03 2 12 0
1 2018-11-01 2018-11-03 2 13 0

Finally, to acquire the survival function data, I group by day (days following start_date) and average the value in active. This provides us with the proportion of users who are active X days after start_date.

day user_count percent_active
0 5 1.0
1 5 0.8
2 5 0.6
3 5 0.6
4 5 0.6
5 5 0.6
6 5 0.4
7 5 0.4
8 5 0.4
9 5 0.4
10 5 0.2
11 5 0.2
12 5 0.2
13 5 0.2

# Looking Towards the Future of Automated Machine-learning

I recently gave a presentation at Venture Cafe describing how I see automation changing python, machine-learning workflows in the near future.

In this post, I highlight the presentation’s main points. You can find the slides here.

From Ray Kurzweil’s excitement about a technological singularity to Elon Musk’s warnings about an A.I. Apocalypse, automated machine-learning evokes strong feelings. Neither of these futures will be true in the near-term, but where will automation fit in your machine-learning workflows?

Our existing machine-learning workflows might look a little like the following (please forgive the drastic oversimplification of a purely sequential progression across stages!).

Where does automation exist in this workflow? Where can automation improve this workflow?

Not all these stages are within the scope of machine-learning. For instance, while you should automate gathering data, I view this as a data engineering problem. In the image below, I depict the stages that I consider ripe for automation, and the stages I consider wrong for automation. For example, data cleaning is too idiosyncratic to each dataset for true automation. I “X” out model evaluation as wrong for automation. In retrospect, I believe this is a great place for automation, but I don’t know of any existing python packages handling it.

I depict feature engineering and model selection as the most promising areas for automation. I consider feature engineering as the stage where advances in automation can have the largest impact on your model performance. In the presentation, I include a strong quote from a Quora user saying that hyper-parameter tuning (a part of model selection) “hardly matters at all.” I agree with the sentiment of this quote, but it’s not true. Choosing roughly the correct hyper-parameter values is VERY important, and choosing the very best hyper-parameter values can be equally important depending on how your model is used. I highlight feature engineering over model selection because automated model selection is largely solved. For example grid-search automates model selection. It’s not a fast solution, but given infinite time, it will find the best hyper-parameter values!

There are many python libraries automating these parts of the workflow. I highlight three libraries that automate feature engineering.

The first is teapot. Teapot (more or less) takes all the different operations and models available in scikit-learn, and allows you to assemble these operations into a pipeline. Many of these operations (e.g., PCA) are forms of feature engineering. Teapot measures which operations lead to the best model performance. Because Teapot enables users to assemble SO MANY different operations, it utilizes a genetic search algorithm to search through the different possibilities more efficiently than grid-search would.

The second is auto_ml. In auto_ml users simply pass a dataset to the software and it will do model selection and hyper-parameter tuning for you. Users can also ask the software to train a deep learning model that will learn new features from your dataset. The authors claim this approach can improve model accuracy by 5%.

The third is feature tools. Feature Tools is the piece of automation software whose future I am most excited about. I find this software exciting because users can feed it pre-aggregated data. Most machine-learning models expect that for each value of the response variable, you supply a vector of explanatory variables. This is an example of aggregated data. Teapot and auto_ml both expect users to supply aggregated data. Lots of important information is lost in the aggregation process, and allowing automation to thoroughly explore different aggregations will lead to predictive features that we would not have created otherwise (any many believe this is why deep learning is so effective). Feature tools explores different aggregations all while creating easily interpreted variables (in contrast to deep learning). While I am excited about the future of feature tools, it is a new piece of software and has a ways to go before I use it in my workflows. Like most automation machine-learning software it’s very slow/resource intensive. Also, the software is not very intuitive. That said, I created a binder notebook demoing feature tools, so check it out yourself!

We should always keep in mind the possible dangers of automation and machine-learning. Removing humans from decisions accentuates biases baked into data and algorithms. These accentuated biases can have dangerous effects. We should carefully choose which decisions we’re comfortable automating and what safeguards to build around these decisions. Check out Cathy O’Neil’s amazing Weapons for Math Destruction for an excellent treatment of the topic.

This post makes no attempt to give an exhaustive view of automated machine-learning. This is my single view point on where I think automated machine-learning can have an impact on your python workflows in the near-term. For a more thorough view of automated machine-learning, check out this presentation by Randy Olson (the creator of teapot).

# Python Aggregate UDFs in PySpark

PySpark has a great set of aggregate functions (e.g., count, countDistinct, min, max, avg, sum), but these are not enough for all cases (particularly if you’re trying to avoid costly Shuffle operations).

PySpark currently has pandas_udfs, which can create custom aggregators, but you can only “apply” one pandas_udf at a time. If you want to use more than one, you’ll have to preform multiple groupBys…and there goes avoiding those shuffles.

In this post I describe a little hack which enables you to create simple python UDFs which act on aggregated data (this functionality is only supposed to exist in Scala!).

id value
1 'a'
1 'b'
1 'b'
2 'c'

I use collect_list to bring all data from a given group into a single row. I print the output of this operation below.

id value_list
1 ['a', 'b', 'b']
2 ['c']

I then create a UDF which will count all the occurences of the letter ‘a’ in these lists (this can be easily done without a UDF but you get the point). This UDF wraps around collect_list, so it acts on the output of collect_list.

id a_count
1 1
2 0

There we go! A UDF that acts on aggregated data! Next, I show the power of this approach when combined with when which let’s us control which data enters F.collect_list.

First, let’s create a dataframe with an extra column.

id value1 value2
1 1 'a'
1 2 'a'
1 1 'b'
1 2 'b'
2 1 'c'

Notice, how I included a when in the collect_list. Note that the UDF still wraps around collect_list.

id a_count
1 1
2 0

There we go! Hope you find this info helpful!

# Custom Email Alerts in Airflow

Apache Airflow is great for coordinating automated jobs, and it provides a simple interface for sending email alerts when these jobs fail. Typically, one can request these emails by setting email_on_failure to True in your operators.

These email alerts work great, but I wanted to include additional links in them (I wanted to include a link to my spark cluster which can be grabbed from the Databricks Operator). Here’s how I created a custom email alert on job failure.

First, I set email_on_failure to False and use the operators’s on_failure_callback. I give on_failure_callback the function described below.

send_email is a function imported from Airflow. contextDict is a dictionary given to the callback function on error. Importantly, contextDict contains lots of relevant information. This includes the Task Instance (key=’ti’) and Operator Instance (key=’task’) associated with your error. I was able to use the Operator Instance, to grab the relevant cluster’s address and I included this address in my email (this exact code is not present here).

To use the notify_email, I set on_failure_callback equal to notify_email.

I write out a short example airflow dag below.

Note where set on_failure_callback equal to notify_email in the PythonOperator.

Hope you find this helpful! Don’t hesitate to reach out if you have a question.

# Aggregating Sparse and Dense Vectors in PySpark

Many (if not all of) PySpark’s machine learning algorithms require the input data is concatenated into a single column (using the vector assembler command). This is all well and good, but applying non-machine learning algorithms (e.g., any aggregations) to data in this format can be a real pain. Here, I describe how to aggregate (average in this case) data in sparse and dense vectors.

I start by importing the necessary libraries and creating a spark dataframe, which includes a column of sparse vectors. Note that I am using ml.linalg SparseVector and not the SparseVector from mllib. This makes a big difference!

row_num features
1 (10,[1,2,3,4,5],[1.0, 1.0, 2.0, 1.0, 3.0])
2 (10,[9],[100.0])
3 (10,[1],[1.0])

Next, I write a udf, which changes the sparse vector into a dense vector and then changes the dense vector into a python list. The python list is then turned into a spark array when it comes out of the udf.

row_num features features_array
1 (10,[1,2,3,4,5],[1.0, 1.0, 2.0, 1.0, 3.0]) [0.0, 1.0, 1.0, 2.0, 1.0, 3.0, 0.0, 0.0, 0.0, 0.0]
2 (10,[9],[100.0]) [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 100.0]
3 (10,[1],[1.0]) [0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]

Now that the data is in a PySpark array, we can apply the desired PySpark aggregation to each item in the array.

averages
[0.0, 0.66667, 0.33333, 0.66667, 0.33333, 1.0, 0.0, 0.0, 0.0, 33.33333]

Now, let’s run through the same exercise with dense vectors. We start by creating a spark dataframe with a column of dense vectors.

row_num features
1 [0.0, 1.0, 1.0, 2.0, 1.0, 3.0, 0.0, 0.0, 0.0, 0.0]
2 [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 100.0]
3 [0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]

Next, we create another PySpark udf which changes the dense vector into a PySpark array.

row_num features features_array
1 [0.0, 1.0, 1.0, 2.0, 1.0, 3.0, 0.0, 0.0, 0.0, 0.0] [0.0, 1.0, 1.0, 2.0, 1.0, 3.0, 0.0, 0.0, 0.0, 0.0]
2 [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 100.0] [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 100.0]
3 [0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0] [0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]

Finally, we can use our standard PySpark aggregators to each item in the PySpark array.

averages
[0.0, 0.66667, 0.33333, 0.66667, 0.33333, 1.0, 0.0, 0.0, 0.0, 33.33333]

There we go! Hope you find this info helpful!

# Integrating Apache Airflow and Databricks

Cron is great for automation, but when tasks begin to rely on each other (task C can only run after both tasks A and B finish) cron does not do the trick.

Apache Airflow is open source software (from airbnb) designed to handle the relationship between tasks. I recently setup an airflow server which coordinates automated jobs on databricks (great software for coordinating spark clusters). Connecting databricks and airflow ended up being a little trickier than it should have been, so I am writing this blog post as a resource to anyone else who attempts to do the same in the future.

For the most part I followed this tutorial from A-R-G-O when setting up airflow. Databricks also has a decent tutorial on setting up airflow. The difficulty here is that the airflow software for talking to databricks clusters (DatabricksSubmitRunOperator) was not introduced into airflow until version 1.9 and the A-R-G-O tutorial uses airflow 1.8.

Airflow 1.9 uses Celery version >= 4.0 (I ended up using Celery version 4.1.1). Airflow 1.8 requires Celery < 4.0. In fact, the A-R-G-O tutorial notes that using Celery >= 4.0 will result in the error:

I can attest that this is true! If you use airflow 1.9 with Celery < 4.0, everything might appear to work, but airflow will randomly stop scheduling jobs after awhile (check the airflow-scheduler logs if you run into this). You need to use Celery >= 4.0! Preventing the Wrong destination error is easy, but the fix is hard to find (hence why I wrote this post).

After much ado, here’s the fix! If you follow the A-R-G-O tutorial, install airflow 1.9, celery >=4.0 AND set broker_url in airflow.cfg as follows:

Note that compared to the A-R-G-O tutorial, I am just adding “py” in front of amqp. Easy!

# Random Weekly Reminders

I constantly use google calendar to schedule reminder emails, but I want some of my reminders to be stochastic!

Google calendar wants all their events to occur on a regular basis (e.g., every Sunday), but I might want a weekly reminder email which occurs on a random day each the week.

I wrote a quick set of python scripts which handle this situation.

The script find_days.py chooses a random day each week (over a month) on which a reminder email should be sent. These dates are piped to a text file (dates.txt). The script send_email.py reads this text file and sends a reminder email to me if the current date matches one of the dates in dates.txt.

I use cron to automatically run these scripts on a regular basis. Cron runs find_days.py on the first of each month and runs send_email.py every day. I copied my cron script as cron_job.txt.

I use mailutils and postfix to send the reminder emails from the machine. Check out this tutorial for how to set up a send only mail server. The trickiest part of this process was repeatedly telling gmail that my emails were not spam.

Now I receive my weekly reminder on an unknown date so I can act spontaneous!

# Regression of a Proportion in Python

I frequently predict proportions (e.g., proportion of year during which a customer is active). This is a regression task because the dependent variables is a float, but the dependent variable is bound between the 0 and 1. Googling around, I had a hard time finding the a good way to model this situation, so I’ve written here what I think is the most straight forward solution.

I am guessing there’s a better way to do this with MCMC, so please comment below if you know a better way.

Let’s get started by importing some libraries for making random data.

Create random regression data.

Shrink down the dependent variable so it’s bound between 0 and 1.

Make a quick plot to confirm that the data is bound between 0 and 1.

All the data here is fake which worries me, but beggars can’t be choosers and this is just a quick example.

Below, I apply a plain GLM to the data. This is what you would expect if you treated this as a plain regression problem

Here’s the actual values plotted (x-axis) against the predicted values (y-axis). The model does a decent job, but check out the values on the y-axis - the linear model predicts negative values!

Obviously the linear model above isn’t correctly modeling this data since it’s guessing values that are impossible.

I followed this tutorial which recommends using a GLM with a logit link and the binomial family. Checking out the statsmodels module reference, we can see the default link for the binomial family is logit.

Below I apply a GLM with a logit link and the binomial family to the data.

Here’s the actual data (x-axis) plotted against teh predicted data. You can see the fit is much better!

CPython 3.6.3
IPython 6.1.0

numpy 1.13.3
matplotlib 2.0.2
sklearn 0.19.1
seaborn 0.8.0
statsmodels 0.8.0

compiler   : GCC 7.2.0
system     : Linux
release    : 4.13.0-38-generic
machine    : x86_64
processor  : x86_64
CPU cores  : 4
interpreter: 64bit