Measure the Strength of Association Between Two Categorical Variables: Mosaic Plot and Chi-Square Test

In a Data Science project it’s really important to get the more insights out of your data. There is a specific phase, the first one in the project, that has the data analysis as goal: the Data Exploration phase.

Among other kinds of analysis, one of the most interesting is the bi-variate one, that finds out the relationship between two variables. If the two variables are categorical, the most common plot used to analyze their relationship is the mosaic plot. At first sight it may appear a little bit confusing. People not aware of some statistical concepts can miss important information this plot can give us. So, we’ll go a little bit deeper in these concepts.

Read the rest of the article here.

Python for SQL Server Specialists Part 4: Python and SQL Server

Python for SQL Server Specialists Part 4: Python and SQL Server

In the last article in this series about Python for SQL Server specialists, you are going to learn how to use SQL Server Python libraries in SQL Server. You can use two scalable libraries, the revoscalepy and microsoftml libraries, which correspond to equivalent R libraries.

SQL Server and ML Integration

With SQL Server 2016 and 2017, you get a highly scalable Machine Learning (ML) engine. Not every function and algorithm is rewritten as a scalable one. Nevertheless, you will probably find the one you need for your analysis of a big dataset. You can store a Python or R data mining or machine learning model in a SQL Server table and use it for predictions on new data. You can even store graphs in a binary column and use it in SQL Server Reporting Services (SSRS) reports. Finally, ML support is not limited to SQL Server only. You can use R code also in Power BI Desktop and Power BI Service, where we at this time (April 2018) still waiting for Python support. However, you can already use both languages, Python and R, in Azure Machine Learning (Azure ML) experiments.

Microsoft provides the highly scalable ML engine in two flavors:
• ML Services (In-Database): This is the installation that integrates ML into SQL Server. It includes a database service that runs outside the SQL Server Database Engine and provides a communication channel between the Database Engine and R runtime. You install it with SQL Server setup. The ML engine includes the open source R and Python components and in addition a set of scalable R and Python packages.
• Microsoft ML Server: This is a standalone ML server with the same open and scalable packages that run on multiple platforms.

Some of the scalable packages shipped with SQL Server R Services are:
• RevoScaleR (for R) and revoscalepy (for Python): This is a set of parallelized scalable functions for processing data, data overview and preliminary analysis, and machine learning models. The procedures in this package can work with chunks of data at a time, so they do not need to load all of the data in memory immediately.
• MicrosoftML (for R) and microsoftml (for Python): This is a package from December 2016, with many additional scalable machine-learning algorithms implemented.
The following figure shows how the communication process between SQL Server and ML engine works:

The components involved and their communications are as follows:
• In SQL Server Database Engine, you run R or Python script with the sys.sp_execute_external_script system stored procedure. SQL Server sends the request to the Launchpad service, a new service that supports the execution of external scripts.
• The Launchpad service starts the launcher appropriate for the language of your script, either the RLauncher.dll or the PyLauncher.dll, and therefore you can launch an external script from SQL Server using the R or Python language. You can see that the infrastructure is prepared to enable the execution of scripts in additional programming languages.
• The RLauncher or the PyLauncher starts RTerm.exe, the R terminal application for executing R scripts, or Python.exe, the Python terminal application.
• The terminal application in any of the two languages sends the script to BxlServer. This is a new executable used for communication between SQL Server and the ML engine. The scalable ML functions are implemented in this executable as well.
• The BxlServer uses SQL Satellite, a new extensibility API that provides a fast data transfer between SQL Server and external runtime. Again, currently the R and the Python runtimes are supported.

Using the Scalable Functions

Time to start using the scalable function. I will show you how to use some of the functions from the revoscalepy package, which is part of the imports at the beginning of the following code. The code defines the connection string to my local SQL Server, the AdventureWorksDW2016 database. Remember, the dbo.vTargetMail view comes from this database. Also note that the RUser used to connect to SQL Server needs permission to use the sys.sp_execute_external_script procedure. The code also defines the chunk size of 1,000 rows.

import numpy as np
import pandas as pd
import pyodbc
import revoscalepy as rp
# Connection string and chunk size
sqlConnStr = """Driver=SQL Server;Server=SQL2017EIM;
chunkSize = 1000;


The next code defines the query to read the data from SQL Server. Note that an ODBC connection is not needed. The RxSqlServerData() function generates a SQL Server data source object. You can think of it as a proxy object to the SQL Server rowset, which is the result of the query. The data itself is not stored in the memory.

# Define the query
TMquery = """SELECT EnglishOccupation AS Occupation,
YearlyIncome, Age, TotalChildren,
NumberCarsOwned, BikeBuyer
FROM dbo.vTargetMail"""
# Only creates the data source object and does not populate it
sqlTM = rp.RxSqlServerData(sql_query = TMquery,
connection_string = sqlConnStr,
string_as_factors = True,
rows_per_read = chunkSize);


Of course, you can also load the data in an in-memory data frame, like the following code does.

# Import the data in memory
TMSQL = rp.rx_import(input_data = sqlTM, report_progress = 3);


With the rx_get_info() function, you can get some basic information about both, the proxy object and the in-memory data frame.

# Get info about the data source and memory data


Here are the results. You can see that the first object is just a proxy, while the second has rows and columns.

>>> rp.rx_get_info(sqlTM)
Connection string:Driver=SQL Server;Server=SQL2017EIM;
Data Source:SqlServer

>>> rp.rx_get_info(TMSQL)
Number of observations:18484.0
Number of variables:6.0


You can use the rx_summary() revoscalepy scalable function to quickly get some descriptive statistics for the data. The function can use both, the proxy object and the in-memory data frame as the source.

# rx_summary() can use the data source or the memory data
sumOut = rp.rx_summary(formula = "~ NumberCarsOwned + Occupation + F(BikeBuyer)",
data = sqlTM)
sumOut = rp.rx_summary(formula = "~ NumberCarsOwned + Occupation + F(BikeBuyer)",
data = TMSQL)


For the sake of brevity, I am showing only one result here.

Summary Statistics Results for: ~ NumberCarsOwned + Occupation + F(BikeBuyer)

Number of valid observations: 18484.0

Name Mean StdDev Min Max ValidObs MissingObs
0 NumberCarsOwned 1.502705 1.138394 0.0 4.0 18484.0 0.0

Category Counts for Occupation
Number of categories: 5

Professional 5520.0
Management 3075.0
Skilled Manual 4577.0
Clerical 2928.0
Manual 2384.0

Category Counts for F(BikeBuyer)
Number of categories: 2

1 9352.0
2 9132.0


The functions that do not come from the revoscalepy package cannot use the proxy object. For example, the following code uses the pandas crosstab() function to do a crosstabulation on the data from the in-memory data frame and from the proxy object. Note that the code that tries to use the proxy object produces an error.

# Pandas crosstab can use the data, but can't use the data source object
pd.crosstab(TMSQL.NumberCarsOwned, TMSQL.TotalChildren)
# The next row produces an error
pd.crosstab(sqlTM.NumberCarsOwned, sqlTM.TotalChildren)
# AttributeError: 'RxSqlServerData' object has no attribute 'NumberCarsOwned'


The next step is initializing and training a linear regression model, using number of cars owned as the target variable, and income, age and number of children as input variables. Please note the syntax of the rx_lin_mod() function – it actually uses R syntax for the function parameters. This syntax might be simpler for you if you already use R; however, it might look a bit weird to pure Python developers.

# Create a linear regression model
linmod = rp.rx_lin_mod(
"NumberCarsOwned ~ YearlyIncome + Age + TotalChildren",
data = sqlTM)


Finally, the following code makes the predictions on the in-memory data frame and shows the first ten rows of those predictions.

# Predictions on the memory data frame
predmod = rp.rx_predict(linmod, data=TMSQL, output_data = TMSQL)


Note that the RevoScaleR package for R ( includes many functions more than the revoscalepy package ( currently supports for Python.

Executing Python Code in SQL Server

Now you need to switch to SQL Server Management Studio (SSMS). You will use Python inside T-SQL code. If you did not configure your SQL Server to allow external scripts, you have to do it now.

-- Configure SQL Server to allow external scripts
USE master;
EXEC sys.sp_configure 'show advanced options', 1;
EXEC sys.sp_configure 'external scripts enabled', 1;
-- Restart SQL Server
-- Check the configuration
EXEC sys.sp_configure;


You can immediately check whether you can run Python code with the sys.sp_execute_external_script procedure. The following code returns a 1×1 table, with value 1 in the single cell.

-- Check whether Python code can run
EXECUTE sys.sp_execute_external_script
@language =N'Python',
OutputDataSet = InputDataSet
print("Input data is: \n", InputDataSet)
@input_data_1 = N'SELECT 1 as col';


And finally, here is the big code that runs Python to create the same linear regression model as before, however this time within SQL Server. In the result, you get the actual data with the predicted number of cars.

-- Create a model and use it inside SQL Server
USE AdventureWorksDW2016;
EXECUTE sys.sp_execute_external_script
@language =N'Python',
from revoscalepy import rx_lin_mod, rx_predict
import pandas as pd
linmod = rx_lin_mod(
"NumberCarsOwned ~ YearlyIncome + Age + TotalChildren",
data = InputDataSet)
predmod = rx_predict(linmod, data = InputDataSet, output_data = InputDataSet)
OutputDataSet = predmod
@input_data_1 = N'
SELECT CustomerKey, CAST(Age AS INT) AS Age,
CAST(YearlyIncome AS INT) AS YearlyIncome,
TotalChildren, NumberCarsOwned
FROM dbo.vTargetMail;'
"CustomerKey" INT NOT NULL,
"YearlyIncome" INT NOT NULL,
"TotalChildren" INT NOT NULL,
"NumberCarsOwned" INT NOT NULL,
"NumberCarsOwned_Pred" FLOAT NULL));


Before finishing this article, let me point out casts in the input SELECT statement. In comparison to SQL Server, Python supports a limited number of data types. Some conversions between SQL Server data types can be done implicitly, other must be done manually. You can read the details about possible implicit conversions at


This article concludes the series about Python for SQL Server specialists. I hope that you enjoyed it, and especially that you gained some new knowledge that you will use soon.

Click here to download the code.

Read the whole series:

Python for SQL Server Specialists Part 1: Introducing Python

Python for SQL Server Specialists Part 2: Working with Data

Python for SQL Server Specialists Part 3: Graphs and Machine Learning


This is just an introduction, for more on starting with data science in SQL Server please refer to the book “Data Science with SQL Server Quick Start Guide” (, where you will learn about tools and methods not just in Python, but also R and T-SQL.

Python for SQL Server Specialists Part 3: Graphs and Machine Learning

Python for SQL Server Specialists Part 3: Graphs and Machine Learning

After learning about Python fundamentals and basics about working with data, it is time to start with more exciting parts of this Python for SQL Server Specialists series.

In this article you will learn about the most important libraries for advanced graphing, namely matplotlib and seaborn, and about the most popular data science library, the scikit-learn library.

Creating Graphs

You will learn how to do graphs with two Python libraries: matplotlib and seaborn. Matplotlib is a mature well-tested, and cross-platform graphics engine. In order to work with it, you need to import it. However, you need also to import an interface to it. Matplotlib is the whole library, and matplotlib.pyplot is a module in matplotlib. Pyplot as the interface to the underlying plotting library that knows how automatically create the figure and axes and other necessary elements to create the desired plot. Seaborn is a visualization library built on matplotlib, adding additional enhanced graphing options, and making work with pandas data frames easy.

Anyway, without further talking, let’s start developing. First, let’s import all necessary packages for this section.

import numpy as np
import pandas as pd
import matplotlib as mpl
import matplotlib.pyplot as plt
import seaborn as sns

The next step is to create sample data. An array of 100 evenly distributed numbers between 0 and 10 is the data for the independent variable, and then the following code creates two dependent variables, one as the sinus of the independent one, and the second as the natural logarithm of the independent one.

# Creating data and functions
x = np.linspace(0.1, 10, 100)
y = np.sin(x)
z = np.log(x)

The following code defines the style to use for the graph and then plots two lines, one for each function. The command is needed to show the graph interactively.

# Basic graph'classic')
plt.plot(x, y)
plt.plot(x, z)

If you execute the code above in Visual Studio 2017, you should get a pop-up window with the desired graph. I am not showing the graph yet; before showing it, I want to make some enhancements and besides showing it also save it to a file. The following code uses the plt.figure() function to create an object that will store the graph. Then for each function defines the line style, line width, line color, and label. The plt.axis() line redefines the axes range. The next three lines define the axes titles and the title of the graph and define font size for the text. The plt.legend() line draws the legend. The last two lines show the graph interactively and save it to a file.

# Enhanced graph
f = plt.figure()
plt.plot(x, y, color = 'blue', linestyle = 'solid',
         linewidth = 4, label = 'sin')
plt.plot(x, z, color = 'red', linestyle = 'dashdot',
         linewidth = 4, label = 'log')
plt.axis([-1, 11, -2, 3.5])
plt.xlabel("X", fontsize = 16)
plt.ylabel("sin(x) & log(x)", fontsize = 16)
plt.title("Enhanced Line Plot", fontsize = 25)
plt.legend(fontsize = 16)

Here is the result of the code above – the first nice graph.

Graphing SQL Server Data

Now it’s time to switch to some more realistic examples. First, let’s import the dbo.vTargetMail data from the AdventureWorksDW2016 demo database in a pandas data frame.

# Connecting and reading the data
import pyodbc
con = pyodbc.connect('DSN=AWDW;UID=RUser;PWD=Pa$$w0rd')
query = """SELECT CustomerKey, 
             TotalChildren, NumberChildrenAtHome,
             HouseOwnerFlag, NumberCarsOwned,
             EnglishEducation as Education,
             YearlyIncome, Age, BikeBuyer
           FROM dbo.vTargetMail;"""
TM = pd.read_sql(query, con)

The next graph you can create is a scatterplot. The following code plots YearlyIncome over Age. Note that the code creates a smaller data frame with first hundred rows only, in order to get less cluttered graph for the demo. Again, for the sake of brevity, I am not showing this graph.

# Scatterplot
TM1 = TM.head(100)
plt.scatter(TM1['Age'], TM1['YearlyIncome'])
plt.xlabel("Age", fontsize = 16)
plt.ylabel("YearlyIncome", fontsize = 16)
plt.title("YearlyIncome over Age", fontsize = 25)

For categorical variables, you usually create bar charts for a quick overview of the distribution. You can do it with the countplot() function from the seaborn package. Let’s try to plot counts for the BikeBuyer variable in the classes of the Education variable.

# Bar chart
sns.countplot(x="Education", hue="BikeBuyer", data=TM);
# Note the wrong order of Education

If you executed the previous code, you should have noticed that the Education variable is not sorted correctly. You need to inform Python about the intrinsic order of a categorial or nominal variable. The following code defines that the Education variable is categorical and then shows the categories,

# Define Education as categorical
TM['Education'] = TM['Education'].astype('category')

In the next step, the code defines the correct order.

 # Proper order 
    ["Partial High School",  
     "High School","Partial College",  
     "Bachelors", "Graduate Degree"], inplace=True)

Now it is time to create the bar chart again. This time, I am also saving it to a file, and showing it here.

# Correct graph
f = plt.figure()
sns.countplot(x="Education", hue="BikeBuyer", data=TM);

So here is the bar chart.

Machine Learning with Scikit-Learn

You can find many different libraries for statistics, data mining and machine learning in Python. Probably the best-known one is the scikit-learn package. It provides most of the commonly used algorithms, and also tools for data preparation and model evaluation.

In scikit-learn, you work with data in a tabular representation by using pandas data frames. The input table (actually a two-dimensional array, not a table in the relational sense) has columns used to train the model. Columns, or attributes, represent some features, and therefore this table is also called the features matrix. There is no prescribed naming convention; however, in most of the Python code, you will note that this features matrix is stored in variable X.

If you have a directed, or supervised algorithm, then you also need the target variable. This is represented as a vector or one-dimensional target array. Commonly, this target array is stored in a variable named y.

Without further hesitation, let’s create some mining models. First, the following code imports all necessary libraries for this section.

# sklear imports
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score
from sklearn.naive_bayes import GaussianNB
from sklearn.mixture import GaussianMixture

Next step is to prepare the features matrix and the target array. The following code also checks the shape of both.

# Preparing the data for Naive Bayes
X = TM[['TotalChildren', 'NumberChildrenAtHome',
        'HouseOwnerFlag', 'NumberCarsOwned',
        'YearlyIncome', 'Age']]
y = TM['BikeBuyer']

The first model will be a supervised one, using the Naïve Bayes classification. For testing the accuracy of the model, you need to split the data into the training and the test set. You can use the train_test_split() function from the scikit-learn library for this task.

# Split to the treining and test sets
Xtrain, Xtest, ytrain, ytest = train_test_split(
    X, y, random_state = 0, train_size = 0.7)

Note that the code above puts 70% of the data into the training set and 30% into the test set. The next step is to initialize and train the model with the training data set.

# Initialize and train the model
model = GaussianNB(), ytrain)

That’s it. The model is prepared and trained. You can start using it for making predictions. You can use the test set for predictions and evaluate the model. A very well-known measure is the accuracy. The accuracy is the proportion of the total number of predictions that were correct, defined as the sum of true positive and true negative predictions with the total number of cases predicted. The following code uses the test set for the predictions and then measures the accuracy.

# Predictions and accuracy
ymodel = model.predict(Xtest)
accuracy_score(ytest, ymodel)

You can see that you can do quite advanced analyses with just few lines of code. Let’s make another model, this time an undirected one, using the clustering algorithm. For this one, you don’t need training and test sets, and also not the target array. The only thing you need to prepare is the features matrix.

# Preparing the data for Clustering
X = TM[['TotalChildren', 'NumberChildrenAtHome',
        'HouseOwnerFlag', 'NumberCarsOwned',
        'YearlyIncome', 'Age', 'BikeBuyer']]

Again, you need to initialize and fit the model. Note the following code tries to group cases in two clusters.

# Initialize and train the model
model = GaussianMixture(n_components = 2, covariance_type = 'full')

The predict() function for the clustering model creates the cluster information for each case in the form of a resulting vector. The following code creates this vector and shows it.

# Predictions
ymodel = model.predict(X)

You can add the cluster information to the input feature matrix.

# Add the cluster membership to the source data
X['Cluster'] = ymodel

Now you need to understand the clusters. You can get this understanding graphically. The following code shows how you can use the seaborn lmplot() function to create scatterplot showing the cluster membership of the cases spread over income and age.

# Analyze the clusters
sns.set(font_scale = 3)
lm = sns.lmplot(x = 'YearlyIncome', y = 'Age', 
                hue = 'Cluster',  markers = ['o', 'x'],
                palette = ["orange", "blue"], scatter_kws={"s": 200},
                data = X, fit_reg = False,
                sharex = False, legend = True)
axes = lm.axes
axes[0,0].set_xlim(0, 190000)

The following figure shows the result. You can see that in cluster 0 there are older people with less income, while cluster 1 consists of younger people, with not so distinctively higher income only.

Python for SQL Server 3


Now this was something, right? With couple of lines of code, we succeeded to create very nice graphs and perform quite advanced analyses. We analyzed SQL Server data. However, we did not use neither the scalable Microsoft machine learning libraries nor Python code inside SQL Server yet. Stay tuned – this is left for the last article in this series.

Click here to download the code


Read the whole series:

Python for SQL Server Specialists Part 1: Introducing Python

Python for SQL Server Specialists Part 2: Working with Data

Python for SQL Server Specialists Part 4: Python and SQL Server


This is just an introduction, for more on starting with data science in SQL Server please refer to the book “Data Science with SQL Server Quick Start Guide” (, where you will learn about tools and methods not just in Python, but also R and T-SQL.

Python for SQL Server Specialists Part 2: Working with Data

Python for SQL Server Specialists Part 2: Working with Data

In my previous article, you learned Python fundamentals. I also introduced the basic data structures. You can imagine you need more advanced data structures for analyzing SQL Server data, which comes in tabular format. In Python, there is also the data frame object, like in R. It is defined in the pandas library. You communicate with SQL Server through the pandas data frames. But before getting there, you need first to learn about arrays and other objects from the numpy library.

In this article, you will learn about the objects from the two of the most important Python libraries, namely, as mentioned, numpy and pandas.

A Quick Graph Demo

For a start, let me intrigue you by showing some analytical and graphic capabilities of Python. I am explaining the code just briefly in this section; you will learn more about Python programming in the rest of this article. The following code imports necessary libraries for this demonstration.

import numpy as np
import pandas as pd
import pyodbc
import matplotlib.pyplot as plt

Then we need some data. I am using the data from the AdventureWorksDW2016 demo database, selecting from the dbo.vTargetMail view.

Before reading the data from SQL Server, you need to perform two additional tasks. First, you need to create a login and a database user for the Python session and give the user the permission to read the data. Then, you need to create an ODBC data source name (DSN) that points to this database. In SQL Server Management Studio, connect to your SQL Server, and then in Object Explorer, expand the Security folder. Right-click on the Logins subfolder. Create a new login and a database user in the AdventureWorksDW2016DW database and add this user to the db_datareader role. I created a SQL Server login called RUser with password Pa$$w0rd, and a user with the same name.

After that, I used the ODBC Data Sources tool to create a system DSN called AWDW. I configured the DSN to connect to my local SQL Server with the RUser SQL Server login and appropriate password and change the context to the AdventureWorksDW2016 database. If you’ve successfully finished both steps, you can execute the following Python code to read the data from SQL Server.

# Connecting and reading the data
con = pyodbc.connect('DSN=AWDW;UID=RUser;PWD=Pa$$w0rd')
query = """SELECT CustomerKey, Age,
             YearlyIncome, TotalChildren,
           FROM dbo.vTargetMail;"""
TM = pd.read_sql(query, con)

You can get a quick info about the data you read with the following code:

# Get the information

The code shows you the first five rows and the shape of the data you just read.

Now I can do a quick crosstabulation of the NumberCarsOwned variable by the TotalChildren variable.

# Crosstabulation
obb = pd.crosstab(TM.NumberCarsOwned, TM.TotalChildren)

And here are the first results, a pivot table of the afore mentioned variables.

TotalChildren NumberCarsOwned 0 1 2 3 4 5
0 990 1668 602 419 449 110
1 1747 1523 967 290 286 70
2 1752 162 1876 1047 1064 556
3 384 130 182 157 339 453
4 292 136 152 281 165 235

Let me show the results of the pivot table in a graph. I need just the following two lines:

# Bar chart
obb.plot(kind = 'bar')

You can see the graph in the following figure.

Using the NumPy Data Structures and Methods

NumPy is short for Numerical Python; the library name is numpy. The library provides arrays with much more efficient storage and faster work than basic lists and dictionaries. Unlike basic lists, numpy arrays must have elements of a single data type. The following code imports the numpy package with alias np. Then it checks the version of the library. Then the code creates two one-dimensional arrays from two lists, one with implicit element data type integer, and one with explicit float data type.

# Numpay intro
np.array([1, 2, 3, 4])
np.array([1, 2, 3, 4], dtype = "float32")

You can create multidimensional arrays as well. The following code creates three arrays with three rows and five columns, one filled with zeroes, one with ones, and one with the number pi. Note the functions used for populating arrays.

np.zeros((3, 5), dtype = int)
np.ones((3, 5), dtype = int)
np.full((3, 5), 3.14)

For the sake of brevity, I am showing here only the last array.

array([[ 3.14,  3.14,  3.14,  3.14,  3.14],
       [ 3.14,  3.14,  3.14,  3.14,  3.14],
       [ 3.14,  3.14,  3.14,  3.14,  3.14]])

There are many additional functions that help you populating your arrays. The following code creates four different arrays. The first line creates a linear sequence of numbers between 0 and 20 with step 2. Note that the upper bound 20 is not included in the array. The second line creates uniformly distributed numbers between 0 and 1. The third line creates ten numbers between with standard normal distribution with mean 0 and standard deviation 1. The fourth line creates a 3 by 3 matrix of uniformly distributed integral numbers between 0 and 9.

np.arange(0, 20, 2)
np.random.random((1, 10))
np.random.normal(0, 1, (1, 10))
np.random.randint(0, 10, (3, 3))

Again, for the sake of brevity, I am showing only the last result here.

array([[0, 1, 7],
       [5, 9, 4],
       [5, 5, 6]])

In order to perform some calculations on array elements, you could use mathematical functions and operators from the default Python engine, and operate in loops, element by element. However, the numpy library includes also vectorized version of the functions and operators, which operate on vectors and matrices as a whole, and are much faster than the basic ones. The following code creates a 3 by 3 array of numbers between 0 and 8, shows the array, and then calculates the sinus of each element using a numpy vectorized function.

# Numpy vectorized functions
x = np.arange(0, 9).reshape((3, 3))

And here is the result.

array([[0, 1, 2],
       [3, 4, 5],
       [6, 7, 8]])
array([[ 0.        ,  0.84147098,  0.90929743],
       [ 0.14112001, -0.7568025 , -0.95892427],
       [-0.2794155 ,  0.6569866 ,  0.98935825]])

Numpy includes also vectorized aggregate functions. You can use them for a quick overview of the data in an array, using the descriptive statistics calculations. The following code initializes an array of five sequential numbers:

# Aggregate functions
x = np.arange(1,6)

Here is the array:

array([1, 2, 3, 4, 5])

Now you can calculate the sum and the product of the elements, the minimum and the maximum, the mean and the standard deviation:

# Aggregations
np.min(x), np.max(x)
np.mean(x), np.std(x)

Here are the results:

(15, 120)
 (1, 5)
(3.0, 1.4142135623730951)

In addition, you can also calculate running aggregates, like running sum in the next example.

# Running totals

The running sum result is here.

array([ 1,  3,  6, 10, 15], dtype=int32)

There are many more operations on arrays available in the numpy module. However, I am switching to the next topic in this Python learning tour, to the pandas library.

Organizing Data with Pandas

The pandas library is built on the top of the numpy library. Therefore, in order to use pandas, you need to import numpy first. The pandas library introduces many additional data structures and functions. Let’s start our pandas tour with the panda Series object. This is a one-dimensional array, like numpy array; however, you can define explicitly named index, and refer to that names to retrieve the data, not just to the positional index. Therefore, a pandas Series object already looks like a tuple in the relational model, or a row in a table. The following code imports both packages, numpy and pandas. Then it defines a simple pandas Series, without an explicit index. The series looks like a simple single-dimensional array, and you can refer to elements through the positional index.

# Pandas series
ser1 = pd.Series([1, 2, 3, 4])

Here are the results. I retrieved the second and the third element, position 1 and 2 with zero-based positional index.

1    2
2    3

Now I will create a series with explicitly named index:

# Explicit index
ser1 = pd.Series([1, 2, 3, 4],
                 index = ['a', 'b', 'c', 'd'])

As you could see from the last example, you can refer to elements using the names of the index, which serve like column names in a SQL Server row. And below is the result.

b    2
c    3

Imagine you have multiple series with the same structure stacked vertically. This is the pandas DataFrame object. If you know R, let me tell you that it looks and behaves like the R data frame. You use the pandas DataFrame object to store and analyze tabular data from relational sources, or to export the result to the tabular destinations, like SQL Server. When I read SQL Server data at the beginning of this article, I read the tabular data in a Python data frame. The main difference, compared to a SQL Server table, is that a data frame is a matrix, meaning that you still can refer to the data positionally, and that the order of the data is meaningful and preserved.

Pandas data frame is a very powerful object. You have already seen the graphic capabilities of it in the beginning of this article, when I created a quite nice bar chart. In addition, you can use other data frame methods to get information about your data with help of descriptive statistics. He following code shows how to use the describe() function on the whole data frame to calculate basic descriptive statistics on every single column, and then how to calculate the mean, standard deviation, skewness, and kurtosis, i.e. the first four population moments, for the Age variable.

# Descriptive statistics
TM['Age'].mean(), TM['Age'].std()
TM['Age'].skew(), TM['Age'].kurt()

Let me finish this article with another fancy example. It is quite simple to create even more complex graphs. The following code shows the distribution of the Age variable in histograms and with a kernel density plot.

# Another graph
(TM['Age'] - 20).hist(bins = 25, normed = True,
                      color = 'lightblue')
(TM['Age'] - 20).plot(kind='kde', style='r--', xlim = [0, 80])

You can see the results in the following figure. Note that in the code, I subtracted 20 from the actual age, to get slightly younger population than exists in the demo database.


I guess the first article in this series about Python was a bit dull. Nevertheless, you need to learn the basics before doing anything fancier. I hope that in this article you got more intrigued by the Python capabilities for working on data. In the forthcoming articles, I will go deeper into the graphing capabilities and in data science with Python.


Click here to download the code


Read the whole series:

Python for SQL Server Specialists Part 1: Introducing Python

Python for SQL Server Specialists Part 3: Graphs and Machine Learning

Python for SQL Server Specialists Part 4: Python and SQL Server


This is just an introduction, for more on starting with data science in SQL Server please refer to the book “Data Science with SQL Server Quick Start Guide” (, where you will learn about tools and methods not just in Python, but also R and T-SQL.

Python for SQL Server Specialists Part 1: Introducing Python

Python for SQL Server Specialists Part 1: Introducing Python

Python is one of the most popular programming languages. It is a general purpose high level language. It was created by Guido van Rossum, publicly released in 1991. SQL Server 2016 started to support R, and SQL Server 2017 adds support for Python. Now you can select your preferred language for the data science and even other tasks. R has even more statistical, data mining and machine learning libraries, because it is more widely used in the data science community; however, Python has broader purpose than just data science, and is more readable and might thus be simpler to learn. This is the first of the four articles that introduce Python to SQL Server developers and business intelligence (BI) specialists. This means that the articles are more focused on Python basics and data science, and less on general programming with Python.


Starting with Python

Python is an interpreted language. The philosophy of the language is about the code readability. For example, you use white spaces to delimit code blocks instead of special characters like semicolon or curly brackets. Python supports automatic memory management. It has a dynamic type system. You can use multiple program paradigms in Python, including procedural, object-oriented, and functional programming. You can find Python interpreters for all major operating systems. The reference implementation of Python, namely CPython, is open source software, managed by the non-profit Python Software Foundation. Of course, being open source, also means that there is a reach set of libraries available. Even the standard library is impressive and comprehensive.

In order to start working with Python and R, you need to do some installation. I am not covering general SQL Server and Visual Studio installation, I am just explaining what you need to do to start using Python with SQL Server.


Installing ML Services and VS 2017 for Data Science

You just start SQL Server setup, and then from the Feature Selection page select Database Engine Services, and underneath Machine Learning (ML) Services (In-Database), with Python only, or both languages, R and Python, selected. After that, all you need are client tools, and you can start writing the code. The following figure shows the SQL Server setup Feature Selection page with appropriate features selected.

The next step is installing client tools. Of course, you need SQL Server Management Studio (SSMS). In addition, you might want to install Visual Studio (VS) 2017. You can use either Professional or even free Community edition to develop python (and also R) code.

When installing Visual Studio 2017, be sure to select Python development workload, and then Data science and analytical applications, like the following figure shows. This will install Python language templates, including data science templates, and also R Tools for Visual Studio.

Selecting the Appropriate Python Engine

There you go, you are nearly ready. There is a small trick here. VS 2017 installs also its own Python interpreter. In order to use the scalable, the one installed with SQL Server, the one that enables executing code in the Database Engine context and includes Microsoft scalable libraries, you need to setup an additional Python environment, pointing to the scalable version of the interpreter. The path for this scalable interpreter is, if you installed the default instance of SQL Server, C:\Program Files\Microsoft SQL Server\MSSQL14.MSSQLSERVER\PYTHON_SERVICES\python.exe. You can see how to setup this environment in the following figure.

That’s it. You are ready to start Python programming. Just start a new project and select the Python Application template from the Python folder. You can also explore the Python Machine Learning templates, which include Classifier, Clustering, and Regression projects. If you selected the Python Application template, you should have open the first empty Python script with default name the same as the project name and default extension py, waiting for you to write and interactively execute Python code.

Python Language Basics

Python uses the hash mark for a comment. You can execute Python code in VS 2017 by highlighting the code and simultaneously pressing the Ctrl and Enter keys. You can use either single or double apostrophes for delimiting strings. The first command you will learn is the print() command. Write and execute the following code:

# Hash starts a comment
print("Hello World!")
# This line ignored
print('Printing again.')
print('O"Hara')   # In-line comment

You can observe the code you wrote and the results in the Interactive window, which is by default below the script window, at the bottom left side of the screen.

Python supports all basic mathematical and comparison operators, like you would expect. The following code introduces them. Note that you can combine strings and expressions in a single print() statement.

1 + 2
print("The result of 3 + 20 / 4 is:", 3 + 20 / 4)
10 * 2 - 7
10 % 4
print("Is 7 less or equal to 5?", 7 <= 5)
print("Is 7 greater than 5?", 7 > 5)

The next step is to introduce the variables. Note that Python is case-sensitive. The following code shows how you can assign values to variables and use them for direct computations and as the arguments of a function.

# Integer
a = 2
b = 3
a ** b
# Float
c = 7.0
d = float(5)
print(c, d)

You can define strings inside double or single quotes. This enables you to use single quotes inside a double-quoted string, and vice-versa. You can use the “%?” operator for formatting strings to include variables, where the question mark stands for a single letter denoting the data type of the variable, for example “s” for strings and “d” for numbers. The str.format() method of the string data type allows you to do variable substitutions in a string. Here are some examples.

e = "String 1"
f = 10
print("Let's concatenate string %s and number %d." % (e, f))
four_cb = "String {} {} {} {}"
print(four_cb.format(1, 2, 3, 4))

The result of the previous code is:

Let’s concatenate string String 1 and number 10.
String 1 2 3 4

You can also create multi-lines strings. Just enclose the strings in a pair of three double quotes. You can also use special characters, like tab and line feed. Escape them with a single backslash character plus a letter, for example letter t for a tab and letter n for a line feed.

You can always get interactive help with the help() command. A Python module is a file with default extension .py containing Python definitions and statements. You can import a module into your current script with the import command, and then use the functions and variables defined in that module. Besides modules provided with the installation, you can, of course, develop your own modules, distribute them, and reuse the code.

Using Functions, Branches, and Loops

Like in any serious programming language, you can encapsulate your code inside a function. You define a function with the def name(): command. Functions can use arguments. Functions can also return values. The following code defines two functions, one that has no arguments, and one that has two arguments and returns a value. Note that there is no special ending mark of a function body – the correct indentation tells the Python interpreter where the body of the first function ends, and the definition of the second function starts.

def p_n():
    print("No args...")
def add(a, b):
    return a + b

When you call a function, you can pass parameters as literals, or through variables. You can also do some manipulation with the variables when you pass them as the arguments to a function. The following code shows these possibilities.

# Call with variables and math
a = 10
b = 20
add(a / 5, b / 4)

You can make branches in the flow of your code with the if..elif..else: statement. The following code shows you and example.

a = 10
b = 20
c = 30
if a > b:
    print("a > b")
elif a > c:
    print("a > c")
elif (b < c):
    print("b < c")
    if a < c:
        print("a < c")
    if b in range(10, 30):
        print("b is between a and c")
    print("a is less than b and less than c")

The results of the code are:

b < c
a < c
b is between a and c

The simplest data structure is the list. Python list is a set of comma-separated values (or items) between square brackets. You can use a for or for each loop to iterate over a list. There are many methods supported by a list. For example, you can use the list.append() method to append an element to a list. The following code shows how to create lists and loop over them with the for and foreach loops. Finally, it shows a while loop.

animals = ["cat", "dog", "pig"]
nums = []
for animal in animals:
    print("Animal: ", animal)
for i in range(2, 5):
i = 1
while i <= 10:
    i = i + 1

The last data structure presented in this introduction article is the dictionary. A dictionary is a set of the key – value pairs. You can see an example of a dictionary in the following code.

states = {
    "Oregon": "OR",
    "Florida": "FL",
    "Michigan": "MI"}
for state, abbrev in list(states.items()):
    print("{} is abbreviated {}.".format(state, abbrev))

I mentioned that in Python you can also use object-oriented paradigm. However, going deeper with object-oriented programming with Python is beyond the scope of this and the following articles.


I guess the programming in Python introduced so far was not over exciting. However, you always need to start with basics, and only after you embrace the basics, the exciting part starts. Therefore, don’t miss my next article, when I will introduce the most important data structure for advanced analytics, the data frame structure.


Read the whole series:

Python for SQL Server Specialists Part 2: Working with Data

Python for SQL Server Specialists Part 3: Graphs and Machine Learning

Python for SQL Server Specialists Part 4: Python and SQL Server


This is just an introduction, for more on starting with data science in SQL Server please refer to the book “Data Science with SQL Server Quick Start Guide” (, where you will learn about tools and methods not just in Python, but also R and T-SQL.