Introduction to testing code for data science

Introduction to testing code for data science#

Topic learning objectives#

By the end of this topic, students should be able to:

Fully specify a function and write comprehensive tests against the specification.
Reproducibly generate test data (e.g., data frames, models, plots).
Discuss the observability of unit outputs in data science (e.g., plot objects), and how this should be taken into account when designing software.
Explain why test-driven development (TDD) affords testability
Use exceptions when writing code.
Test if a function is throwing an exception when it should, and that is does not do so when it shouldn’t.
Evaluate test suite quality.

# Limit output of data frame output to 10 lines
options(repr.matrix.max.rows = 10)

Functions#

Functions are individual units of code that perform a specific task.
They are useful for increasing code modularity - this helps with reusability and readability.
Functions can be easily tested to ensure the correct function outputs are given and that they handle user errors in expected ways. These checks help to increase the robustness of your code.

When should you write a function? In practice, when you start re-writing code for the second or third time, it really is time to abstract your code into a function. When does this happen in data science? Think back to your DSCI 100 projects, you may have had redundant code when you:

Repeatedly applied some data cleaning/manipulation to columns in your raw data suring the data wrangling process
Created many similar data visualizations during your exploratory data analysis
Created many similar tidymodels workflows when you were tuning your predictive models

Workflow for writing functions and tests for data science#

How should we get started writing functions and tests for data science? There are many ways one could proceed, however some paths will be more efficient and less error-prone, and more robust than others. Borrowing from software development best practices, one recommended workflow is shown below:

Write the function specifications and documentation - but do not implement the function. This means that you will have an empty function, that specifies and documents what the name of the function is, what arguments it takes, and what it returns.
Plan the test cases and document them. Your tests should assess whether the function works as expected when given correct inputs, as well as that it behaves as expected when given incorrect inputs (e.g., throws an error when the wrong type is given for an argument). For the cases of correct inputs, you will want to test the top, middle and bottom range of these, as well as all possible combinations of argument inputs possible. Also, the test data should be as simple and tractable as possible while still being able to assess your function.
Create test data that is useful for assessing whether your function works as expected. In data science, you likely need to create both the data that you would provide as inputs to your function, as well as the data that you would expect your function to return.
Write the tests to evaluate your function based on the planned test cases and test data.
Implement the function by writing the needed code in the function body to pass the tests.
Iterate between steps 2-5 to improve the test coverage and function.

Example of workflow for writing functions and tests for data science#

Let’s say we want to write a function for a task we repeatedly are performing in our data analysis. For example, summarizing the number of observations in each class. This is a common task performed for almost every classification problem to examine how many classes there are to understand if we are facing a binary or multi-class classification problem, as well as to examine whether there are any class imbalances that we may need to deal with before tuning our models.

1. Write the function specifications and documentation - but do not implement the function:#

The first thing we should do is write the function specifications and documentation. This can effectively represented by an empty function and roxygen2-styled documentation in R as shown below:

#' Count class observations
#'
#' Creates a new data frame with two columns, 
#' listing the classes present in the input data frame,
#' and the number of observations for each class.
#'
#' @param data_frame A data frame or data frame extension (e.g. a tibble).
#' @param class_col unquoted column name of column containing class labels
#'
#' @return A data frame with two columns. 
#'   The first column (named class) lists the classes from the input data frame.
#'   The second column (named count) lists the number of observations for each class from the input data frame.
#'   It will have one row for each class present in input data frame.
#'
#' @export
#'
#' @examples
#' count_classes(mtcars, am)
count_classes <- function(data_frame, class_col) {
  # returns a data frame with two columns: class and count
}

2. Plan the test cases and document them:#

Next, we should plan out our test cases and start to document them. At this point we can sketch out a skeleton for our test cases with code but we are not yet ready to write them, as we first will need to reproducibly create test data that is useful for assessing whether your function works as expected. So considering our function specifications, some kinds of input we might anticipate our function may receive, and correspondingly what it should return is listed in a table below:

Type of input	Function input	Function should return
Correct user input	Data frame with five classes, with 3 observations per class, and an unquoted column name of column containing class labels.	A new data frame with two columns and five rows. One column with the class names, and one with the counts of 3 observations for each class.
Correct user input	Data frame with two classes, with 3 observations per class, and an unquoted column name of column containing class labels.	A new data frame with two columns and two rows. One column with the class names, and one with the counts of 3 for each class.
Correct user input	Data frame with two classes, with 3 observations for one class, 2 observations for another, and an unquoted column name of column containing class labels.	A new data frame with two columns and two rows. One column with the class names, and one with the counts of 3 for one class and 2 for the other class.
Correct user input	Data frame with one class, with 3 observations per class, and an unquoted column name of column containing class labels.	A new data frame with two columns and one row. One column with the class names, and one with the counts of 3 observations for each class.
Correct user input	Data frame with two classes, with multiple observations for one class, and a single observation for the other class. Also an unquoted column name of column containing class labels.	A new data frame with two columns and two rows. One column with the class names, and one with the counts of observations for each class.
Correct user input	An empty data frame with no classes, and no observations per class. Also an unquoted column name of column containing class labels.	A new empty data frame with two columns and 0 rows. One column with the class names, and one with the counts of observations for each class.
Incorrect user input	Data frame with two classes, with multiple observations per class and a vector of class labels as a separate object	An error that reports the value for the `class_col` argument should be an unquoted column name of column containing class labels from the data frame given in to the first argument.
Incorrect user input	A list as the value for the argument to `data_frame`, and the name of one of the list elements as the value for the argument `class_col`	An error that reports the value for the the `data_frame` object should be a data frame or data frame extension (e.g. a tibble)

Next, I sketch out a skeleton for the unit tests. For R, we will use the well maintained and popular testthat R package for writing our tests. For extra resources on testthat beyond what is demonstrated here, we recommend reading:

With testthat we create a test_that statement for each related group of tests for a function. For our example, we will create the four test_that statements shown below:

library(testthat)

test_that("`count_classes` should return a data frame or data frame extension", {
  # tests to be added here
})

test_that("`count_classes` should return a data frame, or data frame extension, 
with the number of rows that corresponds to the number of unique classes 
in the column passed to `class_col`", {
  # tests to be added here
})

test_that("`count_classes` should return a data frame, or data frame extension, 
whose values in the `count` column correspond to the number of observations 
for the group in the `class` column from the original data frame", {
  # tests to be added here
})

test_that("`count_classes` should throw an error when incorrect types 
are passed to `data_frame` and `class_col` arguments", {
  # tests to be added here
})

── Skip (???): `count_classes` should return a data frame or data frame extension ──
Reason: empty test

── Skip (???): `count_classes` should return a data frame, or data frame extension, 
with the number of rows that corresponds to the number of unique classes 
in the column passed to `class_col` ──
Reason: empty test

── Skip (???): `count_classes` should return a data frame, or data frame extension, 
whose values in the `count` column correspond to the number of observations 
for the group in the `class` column from the original data frame ──
Reason: empty test

── Skip (???): `count_classes` should throw an error when incorrect types 
are passed to `data_frame` and `class_col` arguments ──
Reason: empty test

3. Create test data that is useful for assessing whether your function works as expected:#

Now that we have a plan, we can create reproducible test data for that plan! When we do this, we want to keep our data as small and tractable as possible. We want to test things we know the answer to, or can at a minimum calculate by hand. We will use R code to reproducibly create the test data. We will need to do this for the data we will feed in as inputs to our function in the tests, as well as the data we expect our function to return.

# function input for tests
five_classes_3_obs <- data.frame(class_lables = rep(c("class1", "class2", "class3", "class4", "class5"), 3))
two_classes_3_obs <- data.frame(class_lables = rep(c("class1", "class2"), 3))
two_classes_3_and_2_obs <- data.frame(class_lables = c(rep(c("class1", "class2"), 2), "class1"))
two_classes_3_and_1_obs <- data.frame(class_lables = c(rep("class1", 3), "class2"))
one_class_3_obs <- data.frame(class_lables = rep("class1", 3))
empty_df  <- data.frame(class_lables = character(0))
vector_class_labels <- rep(c("class1", "class2"), 3)
two_classes_3_obs_as_list <- list(class_lables = rep(c("class1", "class2"), 3))

# expected function output
five_classes_3_obs_output <- data.frame(class = c("class1", "class2", "class3", "class4", "class5"),
                                        count = rep(3, 5))
two_classes_3_obs_output <- data.frame(class = c("class1", "class2"),
                                count = c(3, 3))
two_classes_3_and_2_obs_output <- data.frame(class = c("class1", "class2"),
                                      count = c(3, 2))
two_classes_3_and_1_obs_output <- data.frame(class = c("class1", "class2"),
                                      count = c(3, 1))
one_class_3_obs_output <- data.frame(class = "class1",
                              count = 3)
empty_df_output <- data.frame(class = character(0),
                              count = numeric(0))

4. Write the tests to evaluate your function based on the planned test cases and test data:#

Now that we have the skeletons for our tests, and our reproducible test data, we can actually write the internals for our tests! We will do this by using expect_* functions from the testthat package. The table below shows some of the most commonly used expect_* functions. However, there are many more that can be found in the testthat expectations reference documentation.

`testthat` test structure:#

test_that("Message to print if test fails", expect_*(...))

Common `expect_*` statements for use with `test_that`#

Is the object equal to a value?#

expect_identical - test two objects for being exactly equal
expect_equal - compare R objects x and y testing ‘near equality’ (can set a tolerance)
expect_equivalent - compare R objects x and y testing ‘near equality’ (can set a tolerance) and does not assess attributes

Does code produce an output/message/warning/error?#

expect_error - tests if an expression throws an error
expect_warning - tests whether an expression outputs a warning
expect_output - tests that print output matches a specified value

Is the object true/false?#

These are fall-back expectations that you can use when none of the other more specific expectations apply. The disadvantage is that you may get a less informative error message.

expect_true - tests if the object returns TRUE
expect_false - tests if the object returns FALSE

test_that("`count_classes` should return a tibble", {
  expect_s3_class(count_classes(two_classes_3_obs, class_lables), "tibble")
})

test_that("`count_classes` should return a data frame, or data frame extension, 
with the number of rows that corresponds to the number of unique classes 
in the column passed to `class_col`", {
  expect_equivalent(count_classes(five_classes_3_obs, class_lables), five_classes_3_obs_output)
  expect_equivalent(count_classes(two_classes_3_obs, class_lables), two_classes_3_obs_output)
  expect_equivalent(count_classes(one_class_3_obs, class_lables), one_class_3_obs_output)
  expect_equivalent(count_classes(empty_df, class_lables), empty_df_output)
})

test_that("`count_classes` should return a data frame, or data frame extension, 
whose values in the `count` column correspond to the number of observations 
for the group in the `class` column from the original data frame", {
  expect_equivalent(count_classes(two_classes_3_and_2_obs, class_lables), two_classes_3_and_2_obs_output)
  expect_equivalent(count_classes(two_classes_3_and_1_obs, class_lables), two_classes_3_and_1_obs_output)
})

test_that("`count_classes` should throw an error when incorrect types 
are passed to `data_frame` and `class_col` arguments", {
  expect_error(count_classes(two_classes_3_obs, vector_class_labels))
  expect_error(count_classes(two_classes_3_obs_as_list, class_lables))
})

── Failure ('<text>:2:3'): `count_classes` should return a tibble ──────────────
count_classes(two_classes_3_obs, class_lables) is not an S3 object

Error in `reporter$stop_if_needed()`:
! Test failed
Traceback:

1. test_that("`count_classes` should return a tibble", {
 .     expect_s3_class(count_classes(two_classes_3_obs, class_lables), 
 .         "tibble")
 . })
2. (function (envir) 
 . {
 .     handlers <- get_handlers(envir)
 .     errors <- list()
 .     for (handler in handlers) {
 .         tryCatch(eval(handler$expr, handler$envir), error = function(e) {
 .             errors[[length(errors) + 1]] <<- e
 .         })
 .     }
 .     attr(envir, "withr_handlers") <- NULL
 .     for (error in errors) {
 .         stop(error)
 .     }
 . })(<environment>)

Wait what??? Most of our tests fail…

Yes, we expect that, we haven’t written our function body yet!

5. Implement the function by writing the needed code in the function body to pass the tests:#

FINALLY!! We can write the function body for our function! And then call our tests to see if they pass!

#' Count class observations
#'
#' Creates a new data frame with two columns, 
#' listing the classes present in the input data frame,
#' and the number of observations for each class.
#'
#' @param data_frame A data frame or data frame extension (e.g. a tibble).
#' @param class_col unquoted column name of column containing class labels
#'
#' @return A data frame with two columns. 
#'   The first column (named class) lists the classes from the input data frame.
#'   The second column (named count) lists the number of observations for each class from the input data frame.
#'   It will have one row for each class present in input data frame.
#'
#' @export
#'
#' @examples
#' count_classes(mtcars, am)
count_classes <- function(data_frame, class_col) {
  if (!is.data.frame(data_frame)) {
    stop("`data_frame` should be a data frame or data frame extension (e.g. a tibble)")
  }

  data_frame |>
    dplyr::group_by({{ class_col }}) |>
    dplyr::summarize(count = dplyr::n()) |>
    dplyr::rename_at(1, ~ "class")
}

Note 1: we recommending using the syntax PACKAGE_NAME::FUNCTION() when writing functions that will be sourced into other files in R to make it explicitly clear what external packages they depend on. This becomes even more important when we create R packages from our functions later.

Note 2: group_by will throw a fairly useful error message of class_col is not found in data_frame, and we we can let group_by handle that error case instead of writing our own exception to throw an error on.

test_that("`count_classes` should return a tibble", {
  expect_s3_class(count_classes(two_classes_3_obs, class_lables), "data.frame")
})

test_that("`count_classes` should return a data frame, or data frame extension, 
with the number of rows that corresponds to the number of unique classes 
in the column passed to `class_col`", {
  expect_equivalent(count_classes(five_classes_3_obs, class_lables), five_classes_3_obs_output)
  expect_equivalent(count_classes(two_classes_3_obs, class_lables), two_classes_3_obs_output)
  expect_equivalent(count_classes(one_class_3_obs, class_lables), one_class_3_obs_output)
  expect_equivalent(count_classes(empty_df, class_lables), empty_df_output)
})

test_that("`count_classes` should return a data frame, or data frame extension, 
whose values in the `count` column correspond to the number of observations 
for the group in the `class` column from the original data frame", {
  expect_equivalent(count_classes(two_classes_3_and_2_obs, class_lables), two_classes_3_and_2_obs_output)
  expect_equivalent(count_classes(two_classes_3_and_1_obs, class_lables), two_classes_3_and_1_obs_output)
})

test_that("`count_classes` should throw an error when incorrect types 
are passed to `data_frame` and `class_col` arguments", {
  expect_error(count_classes(two_classes_3_obs, vector_class_labels))
  expect_error(count_classes(two_classes_3_obs_as_list, class_lables))
})

Test passed 🎊
Test passed 🌈
Test passed 🥇
Test passed 🎉

No message from the test, means the tests passed!

https://media.giphy.com/media/EXFAJtutz5Ig8/giphy.gif

Are we done? For the purposes of this demo, yes! However in practice you would usually cycle through steps 2-5 two-three more times to further improve our tests and and function!

Where do the function and test files go?#

In the workflow above, we skipped over where we should put our functions we will use in our data analyses, as well as where we put the tests for our function, and how we call those tests!

We summarize the answer to these questions below, but highly recommend you explore and test out our demonstration GitHub repository that has a minimal working example of this: ttimbers/demo-tests-ds-analysis

We also have a Python example here: ttimbers/demo-tests-ds-analysis-python

Where does the function go?#

In R, functions should be abstracted to R scripts (plain text files that end in .R) which live in the project’s R directory. Commonly we name the R script with the same name as the function (however, we might choose a more general name if the R script contains many functions).

In the analysis file where we call the function (e.g. eda.ipynb) we need to call source("PATH_TO_FILE_CONTAINING_FUNCTION") before we are able to use the function(s) contained in that R script inside our analysis file.

Where do the tests go?#

The tests for the function should live in tests/testthat/test-FUNCTION_NAME.R, and the code to reproducibly generate helper data for the tests lives in tests/testthat/helper-FUNCTION_NAME.R. The test suite can be run via testthat::test_dir("tests/testthat"). testthat::test_dir("tests/testthat") first runs any files that begin with helper* and then any files that begin with test*.

Convenience functions for setting this up#

Several usethis R package functions can be used to setup the file and directory structure needed for this:

usethis::use_r("FUNCTION_NAME") can be used to create the R script file the function should live in, inside the R directory
usethis::use_testthat() can be used to create the necessary test directories to use testthat’s automated test suite execution function (testthat::test_dir("tests/testthat"))
usethis::use_test("FUNCTION_NAME") can be used to create the test file for each function

Note: tests/testthat/helper-FUNCTION_NAME.R needs to be created manually, as there is no usethis function to automate this.

Reproducibly generating test data#

As highlighted above, where at all possible, we should use code to generate reproducible, simple and tractable helper data for our tests. When using the testthat R package in R to automate the running of the test suite, the convention is to put such code in a file named helper-FUNCTION_NAME.R which should live in the tests/testthat directory.

Common types of test levels in data science#

Unit tests - exercise individual components, usually methods or functions, in isolation. This kind of testing is usually quick to write and the tests incur low maintenance effort since they touch such small parts of the system. They typically ensure that the unit fulfills its contract making test failures more straightforward to understand. This is the kind of tests we wrote for our example for count_classes above.
Integration tests - exercise groups of components to ensure that their contained units interact correctly together. Integration tests touch much larger pieces of the system and are more prone to spurious failure. Since these tests validate many different units in concert, identifying the root-cause of a specific failure can be difficult. In data science, this might be testing whether several functions that call each other, or run in sequence, work as expected (e.g., tests for a tidymodel’s workflow function)

Source: CPSC 310 class notes from Reid Holmes, UBC

Test-driven development (TDD) and testability#

Test-driven development (TDD): a process for developing software, whereby you write test cases that assess the software requirements before you write the software that implements the software requirements/functionality. Hey! We just did that in the workflow above!

Source: Wikipedia

Testability is defined as the degree to which a system or component facilitates the establishment of test objectives and the execution of tests to determine whether those objectives have been achieved.

In order to be successful, a test needs to be able to execute the code you wish to test, in a way that can trigger a defect that will propagate an incorrect result to a program point where it can be checked against the expected behaviour. From this we can derive four high-level properties required for effective test writing and execution. These are controllability, observability, isolateablilty, and automatability.

controllability: the code under test needs to be able to be programmatically controlled
observability: the outcome of the code under test needs to be able to be verified
isolateablilty: the code under test needs to be able to be validated on its own
automatability: the tests should be able to be executed automatically

Source: CPSC 310 & CPSC 410 class notes from Reid Holmes, UBC]

Discussion: Does test-driven development afford testability? How might it do so? Let’s discuss controllability, observability, isolateablilty, and automatability in our case study of test-driven development of count_classes.

Observability of unit outputs in data science#

Observability is defined as the extent to which the response of the code under test (here our functions) to a test can be verified.

Questions we should ask when trying to understand how observable our tests are:

What do we have to do to identify pass/fail?
How expensive is it to do this?
Can we extract the result from the code under test?
Do we know enough to identify pass/fail?

Source: CPSC 410 class notes from Reid Holmes, UBC

These questions are easier to answer and address for code that creates simpler data science objects such as data frames, as in the example above. However, when our code under test does something more complex, such as create a plot object, these questions are harder to answer, or can be answered less fully…

Let’s talk about how we might test code to create plots!

Visual regression testing#

When we use certain data visualization libraries, we might think that we can test all code that generates data visualizations similar to code that generates more traditional data objects, such as data frames.

For example, when we create a scatter plot object with ggplot2, we can easily observe many of it’s values and attributes. We show an example below:

options(repr.plot.width = 4, repr.plot.height = 4)

cars_ggplot_scatter <- ggplot2::ggplot(mtcars, ggplot2::aes(hp, mpg)) + 
    ggplot2::geom_point()

cars_ggplot_scatter

../../_images/91f0ad47ae9a818a7416068f995b583f984f902b9c3b9367fb2d8df5a7fd799f.png

cars_ggplot_scatter$layers[[1]]$geom

<ggproto object: Class GeomPoint, Geom, gg>
    aesthetics: function
    default_aes: uneval
    draw_group: function
    draw_key: function
    draw_layer: function
    draw_panel: function
    extra_params: na.rm
    handle_na: function
    non_missing_aes: size shape colour
    optional_aes: 
    parameters: function
    required_aes: x y
    setup_data: function
    setup_params: function
    use_defaults: function
    super:  <ggproto object: Class Geom, gg>

cars_ggplot_scatter$mapping$x

<quosure>
expr: ^hp
env:  global

And so we could write some tests for a function that created a ggplot2 object like so:

#' scatter2d 
#'
#' A short-cut function for creating 2 dimensional scatterplots via ggplot2.
#'
#' @param data data.frame or tibble
#' @param x unquoted column name to plot on the x-axis from data data.frame or tibble
#' @param y unquoted column name to plot on the y-axis from data data.frame or tibble
#'
#' @return
#' @export
#'
#' @examples
#' scatter2d(mtcars, hp, mpg)
scatter2d <- function(data, x, y) {
    ggplot2::ggplot(data, ggplot2::aes(x = {{x}}, y = {{y}})) +
        ggplot2::geom_point()
}

helper_data <- dplyr::tibble(x_vals = c(2, 4, 6),
                   y_vals = c(2, 4, 6))

helper_plot2d <- scatter2d(helper_data, x_vals, y_vals)

test_that('Plot should use geom_point and map x to x-axis, and y to y-axis.', {
    expect_true("GeomPoint" %in% c(class(helper_plot2d$layers[[1]]$geom)))
    expect_true("x_vals"  == rlang::get_expr(helper_plot2d$mapping$x))
    expect_true("y_vals" == rlang::get_expr(helper_plot2d$mapping$y))
})

Test passed 🥇

However, when we create a similar plot object using base R, we do not get an object back at all…

cars_scatter <- plot(mtcars$hp, mtcars$mpg)

../../_images/fc0ef24301d9417562b6f50cd8ee55f0351c6e3cd2234bbac45c905244980910.png

typeof(cars_scatter)
class(cars_scatter)

'NULL'

So as you can see, testing plot objects can be more challenging. In the cases of several commonly used plotting functions and package in R and Python, the objects created are not rich objects with attributes that can be easily accessed (or accessed at all). Plotting packages likeggplot2 (R) and altair (Python) which do create rich objects with observable values and attributes appear to be exceptions, rather than the rule. Thus, regression testing against an image generated by the plotting function is often the “best we can do”, or because of this history, what is commonly done.

Regression testing is defined as tests that check that recent changes to the code base do not break already implemented features.

Thus, once a desired plot is generated from the plotting function, visual regression tests can be used to ensure that further code refactoring does not change the plot function. Tools for this exist for R in the vdiffr package. Matplotlib uses visual regression testing as well, you can see the docs for examples of this here.

Visual regression testing with `vdiffr`#

Say you have a function that creates a nicely formatted scatter plot using ggplot2, such as the one shown below:

pretty_scatter <- function(.data, x_axis_col, y_axis_col) {
    ggplot2::ggplot(data = .data, 
                    ggplot2::aes(x = {{ x_axis_col }}, y = {{ y_axis_col }})) +
        ggplot2::geom_point(alpha = 0.8, colour = "steelblue", size = 3) +
        ggplot2::theme_bw() +
        ggplot2::theme(text = ggplot2::element_text(size = 14))
}

library(palmerpenguins)
library(ggplot2)
penguins_scatter <- pretty_scatter(penguins, bill_length_mm, bill_depth_mm) + 
    labs(x = "Bill length (mm)", y = "Bill depth (mm)")
penguins_scatter

Warning message:
“Removed 2 rows containing missing values (geom_point).”

../../_images/9fb1c6c436358ab55bbe972c290aecc344ea00b1a59a88ff7ae78eb83b179b7f.png

What is so pretty about this scatter plot? Compared to the default settings of a scatter plot created in ggplot2 this scatter plot has a white instead of grey background, has blue points instead of black, has larger points, and they points have a bit of transparency so you can see some overlapping data points.

Now, say that you want to write tests to make sure that as you further develop and refactor your data visualization code, you do not break it or change the plot (because you have decided you are happy with what it looks like). You can use the vdiffr visual regression testing package to do this. First, you need to abstract the function to an R script that lives in R. For this case, we would create a file called R/pretty_scatter.R that houses the pretty_scatter function shown above.

Then you need to setup a tests directory and test file in which to house your tests that works with the testthat framework (we recommend using usethis::use_testthat() and usethis::use_test("FUNCTION_NAME") to do this).

Finally, add an expectation with vdiffr::expect_doppelganger to your test_that statement:

library(palmerpenguins)
library(ggplot2)
library(vdiffr)
source("../../R/pretty_scatter.R")

penguins_scatter <- pretty_scatter(penguins, bill_length_mm, bill_depth_mm) + 
    labs(x = "Bill length (mm)", y = "Bill depth (mm)")
penguins_scatter

test_that("refactoring our code should not change our plot", {
    expect_doppelganger("pretty scatter", penguins_scatter)
})

Then when you run testthat::test_dir("tests/testthat") to run your test suite for the first time, it will take a snapshot of the figure created in your test for that visualization and save it to tests/testthat/_snaps/EXPECT_DOPPELGANGER_TITLE.svg. Then as you refactor your code, you and run testthat::test_dir("tests/testthat") it will compare a new snapshot of the figure with the existing one. If they differ, the tests will fail. You can then run testthat::snapshot_review() to get an interactive viewer which will let you compare the two data visualizations and allow you to either choose to accept the new snapshot (if you wish to include the changes to the data visualization as part of your code revision and refactoring) or you can stop the app and revert/fix some of your code changes so that the data visualization is not unintentionally changed.

Below we show an example of running testthat::snapshot_review() after we made our tests fail by removing alpha = 0.8 from our pretty_scatter function source:

`vdiffr` demo#

In this GitHub repository, we have created a vdiffr demo based on the case above: ttimbers/vdiffr-demo

To get experience and practice using this, we recommend forking this, and then cloning it so that you can try using this, and building off it locally.

Practice exercises#

Inside RStudio, run testthat::test_dir("tests/testthat") to ensure you can get the tests to pass as they exist in the demo.
Change something about the code in R/pretty_scatter.R that will change what the plot looks like (text size, point colour, type of geom used, etc).
Run testthat::test_dir("tests/testthat") and see if the tests fail. If they do, run testthat::snapshot_review() to explore the differences in the two image snapshots. You may be prompted to install a couple R packages to get this working.
Add another plot function to the project and create a test for it using testthat and vdiffr.

Example tests for altair plots (using object attributes not visual regression testing)#

Here’s a function that creates a scatter plot:

def scatter(df, x_axis, y_axis):
    chart = alt.Chart(df).mark_line().encode(
        alt.X(x_axis + ':Q',
            scale=alt.Scale(zero=False),
              axis=alt.Axis(tickMinStep=1)
        ),
        y=y_axis
    )
    return chart

Here’s some small data to test it on:

small_data = pd.DataFrame({
        'year': np.array([1901, 1902, 1903, 1904, 1905]),
        'measure' : np.array([25, 25, 50, 50, 100])
    })
small_data

Here’s the plot:

small_scatter = scatter(small_data, 'year', 'measure')
small_scatter

Here’s a unit test for the scatter function:

def test_scatter():
    assert small_scatter.encoding.x.field == 'year', 'x_axis should be mapped to the x axis'
    assert small_scatter.encoding.y.field == 'measure', 'y_axis should be mapped to the y axis'
    assert small_scatter.mark == 'line', 'mark should be a line'
    assert small_scatter.encoding.x.scale.zero == False, "x-axis should not start at 0"
    assert small_scatter.encoding.x.axis.tickMinStep == 1, "x-axis small tick step should be 1"

Evaluate test suite quality#

We really want to know how well the code performs its expected tasks and how robustly it responds to unexpected inputs and situations. Answering this complex and qualitative question involves balancing tests value (e.g., protection against regressions and bugs introduced by refactoring) and tests cost (e.g., time for creation and upkeep, time to run the tests). This is hard to do, and takes a lot of time. Also, as hinted in the previous sentence, it is not very quantifiable.

Code coverage is a way to more simply assess code quality - it is useful to answer the question does the test suite validate all the code? This is an assessment that also attempts to be more quantitative and is popular because it can often be calculated automatically. However it does not factor time or test value, so we have to take it with a grain of salt.

Source: CPSC 310 class notes from Reid Holmes, UBC

Black box & white box testing#

Up until now, we have primarily been focused on deriving tests when thinking about the code specifications - what is the functionality of the code, and does it do what we expect. For example, most of the tests we wrote for the count_classes function would fall under the guise of black box testing. After we follow our TDD-inspired workflow for writing functions and tests for data science, we may want to look inside our code and do some testing of how it is implemented. This would fall under the umbrella of white box testing and code coverage can be very helpful here to identify portions of code that is not being tested well, or at all.

Black box testing: is a method of software testing that examines the functionality of an application without peering into its internal structures or workings. This method of test can be applied virtually to every level of software testing: unit, integration, system and acceptance. It is sometimes referred to as specification-based testing.

White box testing: is a method of software testing that tests internal structures or workings of an application, as opposed to its functionality (i.e. black-box testing). In white-box testing an internal perspective of the system, as well as programming skills, are used to design test cases.

Source: Wikipedia

Coverage#

Definition: Proportion of system being executed by the test suite.

usually reported as a percentage: $$Coverage = \frac{covered}{(covered + uncovered)} * 100$$

Coverage metrics:#

There are many, but here are the ones our automated tools in this course will calculate for you:

Metric	Description	Dependent upon control flow
line	lines of code that tests execute	No
branch	number of branches (independent code segments) that tests execute	Yes

What exactly is a branch?#

my_function <- function(x) {
  # Branch 1
  if (condition_met) {
        y = function_a(x)
        z = function_b(y)
  }
  # Branch 2
  else { 
    y = function_b(x)
    z = function_c(y)
  }
  z
}

Adapted from: http://www.ncover.com/blog/code-coverage-metrics-branch-coverage/

How are line and branch coverage different?#

Consider the same example we just saw and the unit test below, let’s manually calculate the coverage using line and branch coverage metrics:

|  my_function <- function(x) {
|  # Branch 1
|    if (x == "pony") {
|      y = function_a(x)
|      z = function_b(y)
|    }
|    # Branch 2
|    else { 
|      y = function_b(x)
|      z = function_c(y)
|    }
|    z
|  }
|
|  test_that("ponies are actually unicorns", {
|    expect_equal(my_function("pony"), ("Actually a unicorn"))
|  })

Note: function definitions are not counted as lines when calculating coverage

Using branch as the coverage metric:#

Branch 1 (lines 3-5) is covered by the test, because if (x == "pony") evaluates to true. Therefore we have one branch covered, and one branch uncovered (lines 8-10), resulting in 50% branch coverage when we plug these numbers into our formula.

$Coverage = \frac{covered}{(covered + uncovered)} * 100$

$Coverage = \frac{1}{(1 + 1)} * 100$

$Coverage = 50\%$

Using line as the coverage metric:#

Lines 3-5 and 12 are executed, and lines 8-10 are not (function definitions are not typically counted in line coverage). There fore we have 4 lines covered, and 3 lines uncovered, resulting in 57% line coverage when we plug these numbers into our formula.

$Coverage = \frac{covered}{(covered + uncovered)} * 100$

$Coverage = \frac{4}{(4 + 3)} * 100$

$Coverage = 57\%$

In this case, both metrics give us relatively similar estimates of code coverage.

But wait, line coverage can be misleading…#

Let’s alter our function and re-calculate line and branch coverage:

|  my_function <- function(x) {
|  # Branch 1
|    if (x == "pony") {
|      y = function_a(x)
|      z = function_b(y)
|      print(z)
|      print("some important message")
|      print("another important message")
|      print("a less important message")
|      print("just makin' stuff up here...")
|      print("out of things to say...")
|      print("how creative can I be...")
|      print("I guess not very...")
|    }
|    # Branch 2
|    else { 
|      y = function_b(x)
|      z = function_c(y)
|    }
|    z
|  }
|
|  test_that("ponies are actually unicorns", {
|    expect_equal(my_function("pony"), ("Actually a unicorn"))
|  })

Using branch as the coverage metric:#

Branch 1 (lines 3-13) is covered by the test, because if (x == "pony") evaluates to true. Therefore we have one branch covered, and one branch uncovered (lines 16-18), resulting in 50% branch coverage when we plug these numbers into our formula.

$Coverage = \frac{covered}{(covered + uncovered)} * 100$

$Coverage = \frac{1}{(1 + 1)} * 100$

$Coverage = 50\%$

Using line as the coverage metric:#

Lines 3-13 and 20 are executed, and lines 16-18 are not (function definitions are not typically counted in line coverage). There fore we have 12 lines covered, and 3 lines uncovered, resulting in 57% line coverage when we plug these numbers into our formula.

$Coverage = \frac{covered}{(covered + uncovered)} * 100$

$Coverage = \frac{12}{(12 + 3)} * 100$

$Coverage = 80\%$

In this case, the different metrics give us very different estimates of code coverage! 🤯

Take home message:#

Use branch coverage when you can, especially if your code uses control flow!

Calculating coverage in R#

We use the covr R package to do this.

Install via R console:

install.packages("covr")

To calculate line coverage and have it show in the viewer pane in RStudio:

covr::report()

Currently covr does not have the functionality to calculate branch coverage. Thus this is up to you in R to calculate this by hand if you really want to know.

Why has this not been implemented? It has been in an now unsupported package (see here), but its implementation was too complicated for others to understand. Automating the calculation of branch coverage is non-trivial, and this is a perfect demonstration of that.

Calculating coverage in Python#

We use the plugin tool pytest-cov to do this.

Install as a package via conda:

conda install pytest-cov

Calculating coverage in Python#

To calculate line coverage and print it to the terminal:

pytest --cov=<directory> 

To calculate line coverage and print it to the terminal:

pytest --cov-branch --cov=<directory>

How does `coverage` in Python actually count line coverage?#

the output from poetry run pytest --cov=src gives a table that looks like this:

---------- coverage: platform darwin, python 3.7.6-final-0 -----------
Name                  Stmts   Miss  Cover
-----------------------------------------
big_abs/big_abs.py        8      2    75%
-----------------------------------------
TOTAL                     9      2    78%

In the column labelled as “Stmts”, coverage is calculating all possible line jumps that could have been executed (these line jumps are sometimes called arcs). This is essentially covered + uncovered lines of code.

Note - this leads coverage to count two statements on one line that are separated by a “;” (e.g., print(“hello”); print(“there”)) as one statement, as well as calculating a single statement that is spread across two lines as one statement.

In the column labelled as “Miss”, this is the number of line jumps not executed by the tests. So our covered lines of code is “Stmts” - “Miss”.

The coverage percentage in this scenario is calculated by: $$Coverage = \frac{(Stmts - Miss)}{Stmts}$$Coverage = \frac{8 - 2}{8} * 100 = 75\%$$

How does `coverage` in Python actually branch coverage?#

the output from poetry run pytest --cov-branch --cov=src gives a table that looks like this:

---------- coverage: platform darwin, python 3.7.6-final-0 -----------
Name                  Stmts   Miss Branch BrPart  Cover
-------------------------------------------------------
big_abs/big_abs.py        8      2      6      3    64%
-------------------------------------------------------
TOTAL                     9      2      6      3    67%

In the column labelled as “Branch”, coverage is actually counting the number of possible jumps from branch points. This is essentially covered + uncovered branches of code.

Note: because coverage is using line jumps to count branches, each if inherently has an else even if its not explicitly written in the code.

In the column labelled as “BrPart”, this is the number of of possible jumps from branch points executed by the tests. This is essentially our covered branches of code.

The branch coverage percentage in this tool is calculated by:

\[Coverage = \frac{(Stmts\:executed + BrPart)}{(Stmts + Branch)}\]

\[Coverage = \frac{((Stmts - Miss) + BrPart)}{(Stmts + Branch)}\]

Note: You can see this formula actually includes both line and branch coverage in this calculation.

So for big_abs/big_abs.py 64% was calculated from: $$Coverage = \frac{((8 - 2) + 3)}{(8 + 6)} * 100 = 64\%$$

Testing in Python resources#

testing in Python with Pytest (from the Python packages book)
Pytest documentation
Testing Software (from the Research Software Engineering with Python book)

Attribution:#

Advanced R by Hadley Wickham
The Tidynomicon by Greg Wilson
CPSC 310 and CPSC 410 class notes by Reid Holmes, UBC

Introduction to testing code for data science

Contents

Introduction to testing code for data science#

Topic learning objectives#

Functions#

Workflow for writing functions and tests for data science#

Example of workflow for writing functions and tests for data science#

1. Write the function specifications and documentation - but do not implement the function:#

2. Plan the test cases and document them:#

3. Create test data that is useful for assessing whether your function works as expected:#

4. Write the tests to evaluate your function based on the planned test cases and test data:#

testthat test structure:#

Common expect_* statements for use with test_that#

Is the object equal to a value?#

Does code produce an output/message/warning/error?#

Is the object true/false?#

5. Implement the function by writing the needed code in the function body to pass the tests:#

Where do the function and test files go?#

Where does the function go?#

Where do the tests go?#

Convenience functions for setting this up#

Reproducibly generating test data#

Common types of test levels in data science#

Test-driven development (TDD) and testability#

Observability of unit outputs in data science#

Visual regression testing#

Visual regression testing with vdiffr#

vdiffr demo#

Practice exercises#

Example tests for altair plots (using object attributes not visual regression testing)#

Evaluate test suite quality#

Black box & white box testing#

Coverage#

Coverage metrics:#

What exactly is a branch?#

How are line and branch coverage different?#

Using branch as the coverage metric:#

Using line as the coverage metric:#

But wait, line coverage can be misleading…#

Using branch as the coverage metric:#

Using line as the coverage metric:#

Take home message:#

Calculating coverage in R#

Calculating coverage in Python#

Calculating coverage in Python#

How does coverage in Python actually count line coverage?#

How does coverage in Python actually branch coverage?#

Testing in Python resources#

Attribution:#

`testthat` test structure:#

Common `expect_*` statements for use with `test_that`#

Visual regression testing with `vdiffr`#

`vdiffr` demo#

How does `coverage` in Python actually count line coverage?#

How does `coverage` in Python actually branch coverage?#