Live-In Testing [LIT]¶
Last Updated: 2016-05-31
The phrase “unit testing” is far too overreaching, and is linked heavily to the theory of Test-Driven Development. It is also intended for a very specific purpose, but is often mis-applied to all built-in tests.
Because we need a testing standard that has no relation to TDD, we have created the Live-In Testing Standard [LIT]. While it is our sole testing standard at MousePaw Media, other projects may choose to use this standard alongside TDD, as it is an unrelated standard.
Philosophy and Limitations¶
The main principle behind the Live-In Testing Standard is that all tests are built into the code being tested, and are compiled alongside of it. Tests are executed while the program is running.
There are several distinct advantages to this:
- Tests can be shipped with the software. Tech support can instruct users to run tests, in order to generate specific and useful debugging information. This information can in turn be used to fix the problem over the phone, or provide the necessary details to file a bug report.
- It is easier to extract useful information from a test output than from a log file.
- Benchmarking and (most) testing can be performed on all supported systems without the need to install special tools or applications.
- a test that must be run before the current test.
- a single test lifetime involves a single setup, any number of runs, and a single teardown.
- starting a single test one time generally has multiple repetitions. One consistency test may, for example, input the same scenario and validate the output 100 times in a single run of the test.
- a single time a test is executed by the developer or user.
- a complete and exact set of variables and conditions for running a test. Tests should control for all variables.
- the actions taken to prepare a test to be run. Called once per lifetime.
- the actions taken once at the end of a test’s lifetime, generally before the test object is deleted.
MousePaw Media’ Golilocks library (a component of PawLIB) is designed specifically for LIT in C++. GoldilocksShell (upcoming in PawLIB 1.1) can be used to quickly create a run-time interface for Goldilocks.
For other languages, the most important thing to remember about implementing an LIT testing framework is that all tests are executed on-demand by the user during run-time.
- If a single test supports multiple scenarios, each scenario should have a
unique code. For example, an Edge Test that
randomly generates six integers for the scenario might also use those six
integers as its scenario code.
- A given test may use any systems for generating the code, so long as the code is a) unique and b) can be used to duplicate the exact scenario in the test.
- The test must be capable of re-running a scenario by an input scenario code. (Necessary for long-distance testing.)
- Scenario codes must be made up of visible and recognizable characters (ideally Latin-alphanumeric with punctuation), and should be no more than 35 characters in length. (Again, necessary for long-distance testing.)
- Particularly large scenarios may need to be randomly generated beforehand, hard-coded, and stored in a lookup table in order to make the scenario code practical.
- If a test relies on the proper functionality of a feature that is not being directly tested and validated, the test for that feature would be considered a dependency. For example, one must test file I/O before testing XML parsing, and XML parsing must be tested before using a test which writes a random XML file as part of scenario generation.
Behavior Tests [B]¶
A Behavior Test tests a single common behavior or feature. These are generally most useful for comparative testing. A single letter may follow the ID to indicate which subversion of the test is being used.
Consistency Tests [C]¶
A Consistency Test repeats one scenario a large number of times within known limits, and ensures the output is valid and consistent. This is designed to test features and catch Heisenbugs.
- Whether a scenario is random or crafted, the test should repeat that scenario exactly multiple times.
- The output of each repetition should be identical to the output of the last repetition.
- Find repetition limits first, and stay well within them.
Developer Tests [D]¶
These are custom tests designed by a specific developer that don’t fit in another category.
Edge Test [E]¶
Edge tests generate random scenarios for testing a given feature.
- Each scenario should be unique and random, as far as is possible and practical.
- Crafted scenarios may be integrated into the test, as long as the test scenario is still random enough to potential end-user situations. (The danger of crafting a scenario is our implicit and innate tendency to avoid triggering problems in the code in our tests.)
- An edge test may either explicitly target the valid creation of non-error response, or error response, but not both in the same test.
- Edge tests should be written to target vulnerable situations, that is, situations that have a higher probability of failing in an unusual scenario.
Fatality Tests [F]¶
Larger tests intended to trigger total program or system crashes. Intended to find the hard limits of the software and its environment. One example of this is to run a Consistency Test until the system’s resources are totally consumed. The results of a Fatality test are generally useful in a) establishing checkpoints and failsafes that prevent the software from crashing or taking out the system, and b) validating comparability with a particular system.
- These should NEVER be run automatically! (Must be run manually.)
- The interface should display an error before running.
- A Fatality Test should not be totally dependent on the program being tested. (Logfile writing is vital.)
- System Fatality Tests should be monitored closely, and designed to crash the system in a manner which allows the test data to be collected.
- Fatality Tests should always generate a scream-and-die situation, so that the cause of the crash can be validated.
Integration Tests [I]¶
Smaller tests that are intended to ensure that connected classes are communicating with one another properly, and that constructors and initialization is functioning properly.
- Works primarily through ping/pong scenarios, which ensure that two classes, objects, or programming structures are able to access each other appropriately.
- Integration tests with OOP situations would generally need to use specialized dynamic allocation of objects, thus allowing the scenario to be fully controlled by the test.
- Integration tests should use the same constructors and initializers that are used in normal program execution. Thus, the necessary code for Integration tests would have to be hard-coded into the regular program structure, though most of the special code would only be triggered by the test itself.
Proposed Tests [P]¶
Any test that is written, but not yet adopted.
Stress Tests [S]¶
Larger tests intended to break fragile code, to make sure they’re stable.
- Should find the breaking point of the targeted code or feature in all conceivable directions. Each direction may be split into a different test, depending on testing and project needs.
- Stress tests should never target a program or system crash. (See Fatality Tests).
- Should run as either an extreme Consistency test or Edge test, but not both in one test.
- Must write to an external log file, otherwise the information cannot be examined.
Test Suite Types¶
A regression suite runs the minimum tests necessary to ensure that all basic program functionality works as expected. A single suite may target only the particular set of code or feature set affected by a change. (For example, we might have a separate Layers Regression Suite.)
- Made up of Consistency, Edge, and Integration tests only.
- A regression suite should be run on every Differential.
- Should only involve tests which can run quickly and automatically.
These are intended to simulate specific use cases. This allows us to ensure that simultaneous use of features isn’t going to create problems.
- Made up of Consistency, Edge, and Integration tests only.
- Should ideally generate semi-crafted scenarios.
- May take longer to run than a Regression Suite, but should only involve tests which can run automatically.
Test Library Structure¶
A test should have a unique identifier in addition to a name. The unique identifier should start with a lowercase ‘t’, and then it should indicate the type and a number. In the case of Developer and Proposed, initials are also required at the end. With Proposed tests, the intended category should precede the number.
- A developer test by Jason C. McDonald (with the number 001).
- Behavior test 51.
- Edge test 105.
- Stress test 68.
- Consistency test 971.
- A consistency test designed by Jason C. McDonald, but not yet officially adopted.
Each section of a project should be given its own “namespace” within test names, to prevent conflicts.
Major sections might be assigned an ID, which can be tacked on the beginning of the test name. For example, in the RATS Game Engine, the following IDs are used.
|A||Anari Graphics System|
|TCE||Trailcrest: Content Engine|
|TUE||Trailcrest: User Engine|
|TWE||Trailcrest: World Engine|
P-tB102 would be a behavior test for PawLIB.
Depending on implementation, all tests for a particular project could be loaded into the test system on-demand.
It may frequently be necessary to further subdivide a project’s tests. The first one or two digits of the test ID can be used to indicate the sector of the project. For example, within PawLIB, we use the following numbers:
Thus, looking again at
P-tB102, that would be a behavior test relating
to data structures. (We actually reserve the second digit for further
10 relates specifically to FlexArray.)
In many cases, tests should start as Proposed (
...-p???). Then they are
added later by the lead developer to the official library. This is to prevent
conflicts when two developers add tests with the same name.
tC679-pJCM would be a proposed test by Jason C. McDonald.
This step may be skipped if a single developer is working alone on a section.
Most tests, with the possible exception of Fatality and some Stress tests, should remain in the code. This way, they can be run from a developer terminal by the end-user, as a component of long-distance technical support.
For example, a tech support agent could ask the user to bring up a developer
Ratscript terminal (which would probably involve entering a unique key),
and then type test
tC971. The results could then be read back to the tech
support agent (i.e.
FAILED: Could not create data structure., at which
point the exact cause of the problem on the user’s computer can be pinpointed.