unitizeR - Reproducible Tests

Brodie Gaslam

Managing State

Reproducibility

R’s emphasis on avoiding side effects generally means that if you run the same R code more than once you can be relatively certain that you will get the same result each time. While this is generally true, there are some exceptions. If you evaluate:

x <- x + 5

on the command line, the result will depend on what the value of x was in the workspace prior to evaluation. Since workspaces are littered with objects from day to day R use tests are better run elsewhere to avoid conflicts with those objects.

There are even more subtle factors that can affect test evaluation. For example, if x is an S3 object, the packages loaded on the search path could affect the result of the command. Global options could also affect the outcome.

Here is a non-exhaustive list of aspects of state that might affect test outcomes:

  1. Workspace / Evaluation Environment.
  2. Random seed.
  3. Working directory.
  4. Search path.
  5. Global options.
  6. Loaded namespaces.
  7. System time.
  8. System variables.
  9. Locale.
  10. etc.

Ideally a unit testing framework would nullify these environmental factors such that the only changes in test evaluation are caused by changes in the code that is being tested. unitizer provides functionality that sets session state to known “clean” values ahead of the evaluation of each test. Currently unitizer attempts to manage the first six aspects of state listed above.

In order to comply with CRAN policies state management is turned off by default.

Batch Evaluation and Deferred Review

unitizer batch processes all the tests when it is first run before it breaks into interactive mode. It does this to:

  1. Display useful summary data (how many tests passed/failed in which sections), which is often helpful to know before beginning to debug.
  2. Allow time consuming process to run unattended so that the interactive test review process is not interrupted by slow tests.

The batch-evaluate-and-review-later creates the need for a mechanism to recreate state for when we review the tests. Imagine trying to figure out why a test failed when all the variables may have been changed by subsequent tests. unitizer will always recreate the state of the variables defined by the test scripts, and can optionally recreate other aspects of state provided that is enabled.

Enabling State Management

You can turn on the “suggested” state management level to manage the first four elements of state listed in the previous section. To do so, use unitize(..., state='suggested') or options(unitizer.state='suggested'). Be sure to read ?unitizerState before you enable this setting as there are cases when state management may not work.

Workspace And Evaluation Environments

Test Environments

In order to allow review of each test in its original evaluation environment, each test is evaluated in a separate environment. Each of these environments has for parent the environment of the previous test. This means that a test has access to all the objects created/used by earlier tests, but not objects created/used by subsequent tests. When a later test “modifies” an existing object, the existing object is not really modified; rather, the test creates a new object of the same name in the child environment which masks the object in the earlier test. This is functionally equivalent to overwriting the object as far as the later test is concerned.

For the most part this environment trickery should be transparent to the user. An exception is the masking of ls and traceback with versions that account for the special nature of the unitizer REPL. Another is that you can not remove an object created in an earlier test with rm (well, it is possible, but the how isn’t documented and you are advised not to attempt it). Here is a more complex exception:

a <- function() b()
NULL                 # Prevent `a` and `b` being part of the same test
b <- function() TRUE
a()

In this case, when we evaluate a() we must step back two environments to find a, but that’s okay. The problem is that once inside a, we must now evaluate b(), but b is defined in a child environment, not a parent environment so R’s object lookup fails. If we remove the NULL this would work, but only because neither the a or b assignments are tests, so both a and b would be assigned to the environment of the a() call (see details on tests vignette).

If you are getting weird “object not found” errors when you run your tests, but the same code does not generate those errors when run directly in the command line, this illusion could be failing you. In those situations, make sure that you assign all the variables necessary right ahead of the test so they will all get stored in the same environment.

The Parent Environment

In the “suggested” state tracking mode unitize will run tests in an environment that has the same parent as .GlobalEnv (UnitizerEnv below):

             .GlobalEnv
                       \
                        +--> package:x --> ... --> Base
                       /
TestEnv --> UnitizerEnv

This means that objects in the global environment / workspace will not affect your tests.

Unfortunately implementing this structure is not trivial because we need to ensure UnitizerEnv stays pointed at the environment just below .GlobalEnv even as tests modify the search path by calling library/attach/detach, etc. To achieve this unitizer traces base::library, base::attach, and base::detach when state tracking is enabled and only when unitizer is running. Any time any of those functions is called, unitizer updates the parent of UnitizerEnv to be the second environment on the search path (i.e. the parent of .GlobalEnv). So, for example, if a test calls library(z), the new search path would look like so:

             .GlobalEnv
                       \
                        +--> package:y --> package:x --> ... --> Base
                       /
TestEnv --> UnitizerEnv

Clearly overriding such fundamental functions such as library / attach / detach is not good form. We recognize this, and try to do the overriding in as lightweight a manner as possible by tracing them only to record the search path while unitizer is evaluating. This should be completely transparent to the user. The untracing is registered to the on.exit of unitize so the functions should get untraced even if unitize fails.

Aside from the issues raised above, this method is not completely robust. Any tests that turn tracing off using tracingState, or themselves trace/untrace any of library / attach / detach will interfere with unitizer. If you must do any of the above you should consider specifying a parent environment for your tests through the state parameter to unitize (see ?unitize).

Some functions that expect to find .GlobalEnv on the search path may not work as expected. For example, setClass uses topenv by default to find an environment to define classes in. When setClass is called at the top level, this normally results in the class being defined in .GlobalEnv, but if .GlobalEnv is not available setClass will attempt to define the class in the first environment on the search path, which will likely be a locked namespace. You can work around this by specifying an environment in calls to setClass.

Package Namespace as Parent Environment

Sometimes it is convenient to use the namespace of a package as the parent environment. This allows you to write tests that use internal package functions without having to resort to :::. You can set the parent evaluation environment with the state argument to unitize / unitize_dir. See ?unitize and ?unitizeState.

If you do use this feature keep in mind that your tests will be directly exposed to the global environment as well since R looks through the search path starting at the global environment after looking in the package namespace and imports (your package code is always exposed to this).

Issues With Reference Objects

For the most part R is a copy-on-modify language, which allows us to employ the trickery described above. There are however “reference” objects that are not copied when they are modified. Notable examples include environments, reference classes, and data.table. Since our trickery requires us to keep copies of each object in different environments as they are modified, it does not work with reference objects since they are not automatically duplicated.

The main consequence of this is that when you are reviewing a test that involves a reference object, the value of that reference object during review will be the value after the last reference modification, which may have been made after the test you are reviewing. The tests will still work as they should, passing if you did not introduce regressions, and failing otherwise. However if you review a failed test you may have a hard time making sense of what happened since the objects you review will may not have the values they had when the test was actually run.

Patchwork Reference Environments

When we review unitizer tests, it is possible to end up in a situation where we wish to update our store by keeping a mix of the new tests as well as some of the old ones. This leads to some complications because in order to faithfully reproduce the environments associated with both the reference and the new tests we would potentially have to store the entire set of environments produced by the test script for both the new and reference tests. Even worse, if we re-run unitizer again, we run the risk of having to store yet another set of environments (the old reference environments, what were new environments but became reference ones on this additional run, and the new environments created by this third run). The problem continues to grow with as each incremental run of the unitizer script potentially creates the need to store yet another set of environments.

As a work-around to this problem unitizer only keeps the environment associated with the actual reference tests you chose to keep (e.g. when you type N at the unitizer prompt when reviewing a failed test). unitizer then grafts that test and its environment to the environment chain from the newly evaluated tests (note that for all tests that pass, we keep the new version of the tests, not the reference one). This means that in future unitizer runs where you examine this same reference test, the other “reference” objects available for inspection may not be from the same evaluation that produced the test. The ls command will highlight which objects are from the same evaluation vs which ones are not (see the discussion on ls).

This is not an ideal outcome, but the compromise was necessary to avoid the possibility of ever increasing unitizer stores. For more details see ?"healEnvs,unitizerItems,unitizer-method".

Clean Search Paths

Description / Implementation

One other way tests can change behavior unexpectedly is if the packages / objects attached to the search path change. A simple example is a test script that relies on package “X”, and the user attached that package at some point during interactive use, but forgot to add the requisite library call to the test script itself. During testing, the scripts will work fine, but at some future date if the test scripts are run again they are likely to fail due to the dependency on the package that is not explicitly loaded in the test scripts.

In the “suggested” state tracking mode unitizer runs on a “trimmed” search path that contains only the packages loaded by in a freshly loaded R session (i.e. the packages between package:base and package:stats; see ?unitizerState). You will need to explicitly load packages that your tests depend on in your test file (e.g. by using library()). unitize will restore the search path to its original state once you complete review.

unitizer also relies on tracing library/attach/detach to implement this feature, so the caveats described above apply equally here. unitizer does not modify the search path itself other than by using library, attach, and detach.

When search path tracking is enabled, unitizer tracks the versions of the packages on the search path. If tests fails and package versions on the search path have changes since the reference test was stored, you will be alerted.

Potential Issues

When unitizer manipulates the search path it restores the original one by using library/attach on any previously detached objects or packages. This generally works fine, but detaching and re-attaching packages is not and cannot be the same as loading a package or attaching an environment for the first time. For example, S3 method registration is not undone when detaching a package, or even unloading its namespace. See discussion in ?detach and in ?unitizerState.

One known problem is the use of devtools::load_all and similar which place a pretend package environment on the search path. Such packages cannot be re-loaded with library so the re-attach process will fail (see #252).

Another issue is attached environments that contain references to themselves, as the tools:rstudio environment attached by Rstudio does. It contains functions that have for environment the tools:rstudio environment. The problem is that once that environment is detached from the search path, those functions no longer have access to the search path. Re-attaching the environment to the search path does not solve the problem because attach attaches a copy of the environment, not the environment itself. This new environment will contain the same objects as the original environment, but all the functions therein will have for environment the original detached environment, not the copy that is attached to the search path.

For the specific tools::rstudio problem we work around the issue by keeping it on the search path even search path tracking is enabled (you can over-ride this by changing search.path.keep, or, if you have environments on your search path with similar properties, add their names to search.path.keep). Other options include re-attaching with parent.env<- instead of attach, but messing with the search path in that way seems to be exactly what R core warns about in ?parent.env:

The replacement function parent.env<- is extremely dangerous as it can be used to destructively change environments in ways that violate assumptions made by the internal C code. It may be removed in the near future.

Global Options

unitizer can track and reset global options. Because many packages set options when their namespaces are attached, implementation of this feature must be coordinated with a careful management of loaded namespaces. For example, we can reasonably easily set options to be what you would expect in a freshly loaded vanilla R session, but if some namespaces as otherwise they would be in a compromised set with their options wiped out.

unitizer can manage search paths and namespaces, but unfortunately some package namespaces cannot be unloaded so options management can be problematic when such packages are involved (one example is data.table). Because of this options management is not enabled in the “suggested” state management mode.

Note that no matter what tests are always run with options(warn=1) and options(error=NULL).

See ?unitizer.opts for more details.

Random Seed

See ?unitizerState.

Working Directory

See ?unitizerState.