unitizer - Interactive R Unit Tests

TL;DR

unitizer simplifies creation, review, and debugging of tests in R. It automatically stores R expressions and the values they produce, so explicit expectations are unnecessary. Every test is easy to write with unitizer because testing and using a function are the same. This encourages non-trivial tests that better represent actual usage.

Tests fail when the value associated with an expression changes. In interactive mode you are dropped directly into the failing test environment so you may debug it.

unitizer is on CRAN:

install.packages('unitizer')

It bakes in a lot of contextual help so you can get started without reading all the documentation. Try the demo to get an idea:

library(unitizer)
demo(unitizer)

Or check out the screencast to see unitizer in action.

Why Another Testing Framework?

Automated Test Formalization

Are you tired of the deparse/dput then copy-paste R objects into test file dance, or do you use testthat::expect_equal_to_reference or other snapshot testing a lot?

With unitizer you interactively review your code as you would when typing it at the R prompt. Then, with a single keystroke, you tell unitizer to store the code, and any values, warnings, or errors it produced, thereby creating a formal regression test.

Streamlined Debugging

Do you wish the nature of a test failure was more immediately obvious?

When tests fail, you are shown a proper diff so you can clearly identify how the test failed:

diff example
diff example

Do you wish that you could start debugging your failed tests without additional set-up work?

unitizer drops you in the test environment so you can debug why the test failed without further ado:

review example
review example

Fast Test Updates

Do you avoid improvements to your functions because that would require painstakingly updating many tests?

The diffs for the failed tests let you immediately confirm only what you intended changed. Then you can update each test with a single keystroke.

How Does unitizer Differ from testthat?

Testing Style

unitizer requires you to review test outputs and confirm they are as expected. testthat requires you to assert what the test outputs should be beforehand. There are trade-offs between these strategies that we illustrate here, first with testthat:

vec <- c(10, -10, 0, .1, Inf, NA)
expect_error(
  log10(letters),
  "Error in log10\\(letters\\) : non-numeric argument to mathematical function\n"
)
expect_equal(log10(vec), c(1, NaN, -Inf, -1, Inf, NA))
expect_warning(log10(vec), "NaNs produced")

And with unitizer:

vec <- c(10, -10, 0, .1, Inf, NA)
log10(letters)                            # input error
log10(vec)                                # succeed with warnings

These two unit test implementations are functionally equivalent. There are benefits to both approaches. In favor of unitizer:

  • Tests are easy to write.
  • Tests with non-trivial outputs are easy to write, which encourages more realistic testing of functionality.
  • Conditions are captured automatically, with no need for special handling.
  • You can immediately review failing tests in an interactive environment.
  • Updating tests when function output legitimately changes is easy.

In favor of testthat:

  • The tests are self documenting; expected results are obvious.
  • Once you write the test you are done; with unitizer you still need to unitize and review the tests.
  • Tests are usually all-plain text, whereas unitizer stores reference values in binary RDSes (see Collaborating with Unitizer).

unitizer is particularly convenient when the tests return complex objects (e.g as lm does) and/or produce conditions. There is no need for complicated assertions involving deparsed objects, or different workflows for snapshots.

Converting testthat tests to unitizer

If you have a stable set of tests it is probably not worth trying to convert them to unitizer unless you expect the code those tests cover to change substantially. If you do decide to convert tests you can use the provided testthat_translate* functions (see ?testthat_translate_file).

unitizer and Packages

The simplest way to use unitizer as part of your package development process is to create a tests/unitizer folder for all your unitizer test scripts. Here is a sample test structure from the demo package:

unitizer.fastlm/         # top level package directory
    R/
    tests/
        run.R            # <- calls `unitize` or `unitize_dir`
        unitizer/
            fastlm.R
            cornerCases.R

And this is what the tests/run.R file would look like

library(unitizer)
unitize("unitizer/fastlm.R")
unitize("unitizer/cornerCases.R")

or equivalently

library(unitizer)
unitize_dir("unitizer")

The path specification for test files should be relative to the tests directory as that is what R CMD check uses. When unitize is run by R CMD check it will run in a non-interactive mode that will succeed only if all tests pass.

You can use any folder name for your tests, but if you use “tests/unitizer” unitize will look for files automatically, so the following work assuming your working directory is a folder within the package:

unitize_dir()          # same as `unitize_dir("unitizer")`
unitize("fast")        # same as `unitize("fastlm.R")`
unitize()              # Will prompt for a file to `unitize`

Remember to include unitizer as a “suggests” package in your DESCRIPTION file.

Things You Should Know About unitizer

unitizer Writes To Your Filesystem

The unitized tests need to be saved someplace, and the default action is to save to the same directory as the test file. You will always be prompted by unitizer before it writes to your file system. See storing unitized tests for implications and alternatives.

Tests Pass If They all.equal Stored Reference Values

Once you have created your first unitizer with unitize, subsequent calls to unitize will compare the old stored value to the new one using all.equal. You can change the comparison function by using unitizer_sect (see tests vignette).

Test Expressions Are Stored Deparsed

This means you need to be careful with expressions that may deparse differently on different machines or with different settings. Unstable deparsing will prevent tests from matching their previously stored evaluations.

For example, in order to avoid round issues with numerics, it is better to use:

num.var <- 14523.2342520  # assignments are not considered tests
test_me(num.var)          # safe

Instead of:

test_me(14523.2342520)    # could be deparsed differently

Similarly issues may arise with non-ASCII characters, so use:

chr <- "hello\u044F"      # assignments are not considered tests
fun_to_test(chr)          # safe

Instead of:

fun_to_test("hello\u044F") # could be deparsed differently

This issue does not affect the result of running the test as that is never deparsed.

Increase Reproducibility with Advanced State Management

unitizer can track and manage many aspects of state to make your tests more reproducible. For example, unitizer can reset your R package search path to what is is found in a fresh R session prior to running tests to avoid conflicts with whatever libraries you happen to have loaded at the time. Your session state is restored when unitizer exits. The following aspects of state can be actively tracked and managed:

  • Search path (including removing the global environment from search path)
  • Random seed
  • Working directory
  • Options
  • Loaded namespaces

State management is turned off by default because it requires tracing some base functions which is against CRAN policy, and generally affects session state in uncommon ways. If you wish to enable this feature use unitize(..., state='suggested') or options(unitizer.state='suggested'). For more details including potential pitfalls see ?unitizerState and the reproducible tests vignette.

Beware of browser/debug/recover

If you enter the interactive browser as e.g. invoked by debug you should exit it by allowing evaluation to complete (e.g. by hitting “c” until control returns to the unitizer prompt). If you instead hit “Q” while in browser mode you will completely exit the unitizer session losing any modifications you made to the tests under review.

Reference Objects

Tests that modify objects by reference are not perfectly suited for use with unitizer. The tests will work fine, but unitizer will only be able to show you the most recent version of the reference object when you review a test, not what it was like when the test was evaluated. This is only an issue with reference objects that are modified (e.g. environments, RC objects, data.table modified with := or set*).

unitizer Is Complex

In order to re-create the feel of the R prompt within unitizer we resorted to a fair bit of trickery. For the most part this should be transparent to the user, but you should be aware it exists in the event something unexpected happens that exposes it. Here is a non-exhaustive list of some of the tricky things we do:

  • Each tests is evaluated in its own environment, a child of the previous test’s environment; because R looks up objects in parent environments it appears that all tests are evaluated in one environment (see interactive environment vignette)
  • We mask some base functions.
  • .Last.value will not work
  • We sink stdout and stderr during test evaluation to capture those streams (see details on tests vignette), though we take care to do so responsibly
  • We parse the test file and extract comments so that they can be attached to the correct test for review
  • The history file is temporary replaced so that your unitizer interactions do not pollute it

Avoid Tests That Require User Input

In particular, you should avoid evaluating tests that invoke debugged functions, or introducing interactivity by using something like options(error=recover), or readline, or some such. Tests will work, but the interaction will be challenging because you will have to do it with stderr and stdout captured…

Avoid running unitize within try / tryCatch Blocks

Doing so will cause unitize to quit if any test expressions throw conditions. See discussion in error handling.

Masked Functions

Some base functions are masked at the unitizer prompt:

  • q and quit are masked to give the user an opportunity to cancel the quit action in case they meant to quit from unitizer instead of R. Use Q to quit from unitizer, as you would from browser.
  • ls is masked with a specialized version for use in unitizer.
  • traceback is masked to report the most recent error in the order presented by the unitizer prompt.

See miscellaneous topics vignette.