On Software Testing and the case for the Common Test Scenario Database

TL;DR

As a participant (and sometimes contributor of ideas) to Test Automation efforts (Strategy, tool selection and implementation), it has been my long-standing opinion that the true gains from automation can be achieved only when Test Scenarios (enshrined as exemplar test data) are a heavily reused resource across the organization. This page is an attempt to "arrive" at that opinion by progressively optimizing the software testing process.

I take the example of testing the hotel search feature and progressively optimize it through 4 stages to illustrate the value of each such optimization; and the value of shared test scenarios in particular.

What is Software testing?

Software Testing is the art and science of validating that a particular piece of software behaves as expected. Since most words in the previous sentence are loaded with meaning, I'll dissect it:

Art and science: Art because some of it is based on the skill of the tester and science because most bits are tractable into well-defined, repeatable procedures with predictable outcomes
Validating: This might involve the atomic step of input-process-output, or a sequence of many such atomic steps (typically known as "following a script"). Each such step (or steps) entails comparing an expected value (or more generally outcome) with a known one
Expected: This implies that there are some requirements that the software is supposed to meet which is an accumulation of the most important (or all) of the comparisons mentioned above.

A Concrete Example - Testing Hotel Search functionality

Any travel website has a hotel search functionality. You enter a destination, start and end dates and number of guests; and the website returns a list of hotels. Each hotel card has some details (name, location, address, picture, reviews, rating, description etc); and the list itself has some details (count, sort order, #pages, etc).

Notice also that there are some implicit inputs (eg: choice of POS, choice of currency) and some implicit outputs (eg: sort order).

To truly test this feature, you need to validate that:

The functionality is right:

The list is as expected (has the expected number of hotels which are in the expected order, paginated at the right place, etc)
Each hotel card is as expected (Has the name as expected, image shows up, description is correct, etc)

The display is right:

Visual characteristics for each data item is correct
No visual elements overlap
And so on

Adding the feature didn't break anything else:

tests on features that already exist pass
tests on features that are in-flight at the same time and targeted to the same release pass

The performance is within acceptable limits

Focusing on functional testing alone for a moment, we see that:

Each combination of values that we input produces a unique output; and
Each such output has details that need to be validated against expected values.

The universe of all things that need to be tested therefore is of the order:

Number of input combinations X Size of output for each input combination

Now, for the hotel search scenario,

Number of input combinations = Product of (Size of the set that an input belongs to) for all inputs.

And

Size of each output = Size of result set details + Size of details for all hotel cards

Note: inputs include both implicit and explicit ones

Using the details mentioned above and ignoring the etc's for now,

Size of inputs for hotel search

Number of inputs = 2 implicit + 3 explicit = 5
Product of sizes
= Size of the set of destinations
X Size of the set of Possible Start dates (and invalid ones)
X Size of the set of Possible End dates (and invalid ones)
X Size of set of number of guests (subset of Integers)
X Size of set of POSs allowed (subset of integers)
X Size of set of Currencies allowed

= 50000 (assumed)
X 365 (assuming a year of allowed booking)
X 365 (assumed similarly)
X 4 (assumed to be max allowed guests per booking)
X 1 (assuming USPOS)
X 1 (assuming USD)

Size of each output

= 3 result set details + N x (7 hotel card details)

Where N = size of result set itself

If N = 100, the Test Universe = 50000 x 365 x 365 x 4 x 1 x 1 x (3 + 100 x 7) = 1E13 Tests

Onto this set of 1e13 Tests, we'll have to add the Regression, UI Validation and Performance Tests as well.

Sidebar for the mathematically inclined

All testing can be considered validation of functions. Given an abstract function y = f(x), testing can be considered validation that for all x in X (the input set), there exists the expected y in Y (the output set)

The domain of the function represents the size of the input and the range of the function represents the size of the output. The Cartesian product is, therefore, the set of tests to be conducted to validate the function.

Obviously, this is a lot of testing to do; and if we're actually able to do all of it, that would be Exhaustive Testing. Also obviously, anything larger than the simplest feature would quickly be intractable due to the combinatorial explosion of tests required; so we apply the "art and science" part and try to pare the problem down a bit. I'll focus on the functional testing here, but most of the concepts apply to UI validation and Performance as well.

Before we start, however, I'll define some factors to evaluate the efficacy of our optimization with:

The number of tests required to be run (or reduced as a result of the optimization)
The cost of running the tests (in terms of time, resources, $, etc)
The overall quality of the product is still within acceptable limits

Optimizing testing - Round 1: Scenario Based Testing

This optimization concept is very simple - reduce each set mentioned above to a representative subset. Each element of this subset would "stand for" or represent an entire class of values within the original set. A success or failure of the chosen value is deemed a success or failure of the entire class.

To continue the Hotel Search example, the inputs could be reduced to the following scenarios (for example):

Destinations	Top 10 destinations 1 non-US destination 3 destinations with "known issues" or special edge cases
Booking Dates	Last minute bookings (1 date in the next 3 days) Planned bookings (4 dates in the next 2 quarters) Peak date bookings (2 National holidays) Weekend bookings (2 weekends in the next 2 months)
Guests	4 is small enough a number to test all combos, but we could still pick a subset, say 1 and 4 only

With this, the input size drops to 13 x 9 x 2 = 234

And the test universe becomes 234 x (3 + 100 x 7 ) = 164,502

That's still a huge number, but 8 orders of magnitude less already! We could optimize this further by reducing the validation on the output values if we want to. Realistically, we can probably get away with checking 4 of the 7 details for the first 25 hotels; and checking the result set details just once throughout. So the test universe reduces to:

234 x ( 25 x 4) + 3 = 23,403

How has this impacted our evaluation factors?

The number of tests required to be run has obviously come down.
The cost of running each of the tests still remains the same; we haven't optimized that yet.
The resultant overall quality depends on the scenarios chosen. If we've chosen well; it should be within acceptable limits.

Note that there are more algorithmic ways of arriving at a subset of tests; Orthogonal Array testing to name one. I'll not elaborate on this further as the optimization is the same in principle - that of reducing the total number of tests required in validating a feature.

Similarly, on the regression testing side of the house, scenario-based reduction of tests can be done by carefully analyzing the areas that changed code is likely to impact; aka Impact Analysis.

Optimizing testing - Round 2: Automation

When you have many features similar to the one depicted above, scale effects come to bear:

The total number of tests to be run is still a large overall number
The cost of regression is a constant despite reducing the regression test suite using impact analysis.

The optimization to counter this is conceptually simple - relegate repeatable tests to a machine so that human cycles can be spent on the unique ones. This is easier said than done in practice; and the quickest way to get started is - surprisingly similar to BDD precepts - Outside In. That is, start at the outermost layer and automate tests at that layer. Work gradually inwards; or even not at all. Automating regression alone can have significant benefits.

One of the biggest issues with automation, however, is that you miss out on the human ingenuity bit. Scripts WILL break if data changes over time, so environments have to be stable; something that the human tester would easily sidestep by validating "outside the box" that the changed data is indeed still valid.

To continue with the Hotel Search example, assuming both the human and the machine take the same time for a test, the gain in human cycles due to various levels of automation are:

Feature	Manual Tests	Human cycles saved with 10% automation	25%	50%	75%
Hotel Search	23403	23404 * .1 = 2340.4	23404 * .25 = 5851	23404 * .5 = 11702	23404 * .75 = 17553.0

Reviewing our factors again, we see that with this additional optimization,

The number of tests required to be run by humans has come down again and the tests run by machines can be run repeatably so.
The cost of running each of the tests has reduced, but the cost of maintaining test environments went up as both manual and automated environments have to be kept running.
The resultant overall quality still depends on the scenarios chosen and our trust in our automated tests. If we've chosen well and trust our scripts; it should still be within the same acceptable limits.

Optimizing testing - Round 3: Component based testing

The cost of maintaining test environments mentioned above is typically the tip of the iceberg. All testing espoused to this point has been strictly end-to-end, ie, the environment has been a live one from the UI all the way to the database (or back-end). There is a non-trivial cost associated with maintaining these environments; and a collateral cost of maintaining scripts (or known data for use by the scripts) as those environments change. Additionally, some kinds of testing may not be possible in live environments. Booking scenarios are typically such tests - contractual obligations or the cost of test bookings may deter such tests from being done in a large scale or at all.

In addition, end-to-end testing forces the entire organization into a train wreck of dependencies. Since all testing is done from the UI, all code must be built and integrated before ANY testing can start. This not only delays testing, it also puts pressure on changes to the inner layers of the application - that code has to be completed WAY IN ADVANCE of the UI code, but cannot validate their output until the UI is done.

Component testing attempts to fix these issues by testing each component

at ITS interface, not at the final user interface. That way, the owners of that component know for sure that they produce valid output for given input; a large live environment need not be maintained; and the validation of a test scenario is spread across multiple component tests which together comprise the larger validation.

Component testing almost always predicates the mocking of dependent components because the cost gains are not realized otherwise. That is, if A -> B -> C is a string of components involved in a particular test scenario, C must be mocked out to test B and B must be mocked out to test A; otherwise we've taken on the additional job of maintaining a separate instances of A,B and C solely for component testing purposes, thereby increasing cost of maintaining environments more than without it.

Component testing also typically requires some means of creating mock data - manual means will not suffice; especially if the request-response payloads are huge object graphs.

The choice, adoption and usage of an organization-wide mocking framework is therefore a non-trivial task and I will not go into the details of how to achieve this. I will, however, analyze the impact of adopting such component testing on the evaluation factors mentioned above.

To continue the Hotel Search example, a hotel search typically involves a string of internal components:

                                                               |------------> GDS

UI -> Business  -> Business  -> Business  ->    Back End-------|------------> CRS1

Dispatcher        Facade     Logic Executor  Abstraction Layer

                                                               |------------> CRS2

(Some details may be missing; sorry. I'm trying to make a larger point here).

Let's take the following Test Scenario:

Input	Expected Output
Top Destination (LAS), Weekend (start: next Friday, ret: next sun), 1 guest	1st page has 25 hotels 1st hotel is Caeasar's Palace @ $100 2nd hotel is MGM @ $125 …and so on

…and break it into component tests:

Component	Given a test script that provides this input	..should provide this output	..using this mock data
UI Layer	LAX, Next fri, Next Sun, 1 guest	1st page has 25 hotels 1st hotel is Caeasar's Palace @ $100 2nd hotel is MGM @ $125 …and so on	Business Dispatcher response containing the 25 hotels
Business Dispatcher/ Facade	LAX, mm/dd/yyyy,mm/dd/yyyy,1 + other Dispatcher required params	Arraylist of Business Layer objects	Executor response containing the 25 hotels
Business Logic Executor	LAX, mm/dd/yyyy,mm/dd/yyyy,1 + other Executor required params	Arraylist of Executor-specific objects	LAX-to-internal-location-object response Back end Abstraction Layer responses
Back end Abstraction Layer	LAX, mm/dd/yyyy,mm/dd/yyyy,1 + other Back end required params	Arraylist of Back end Abstraction Layer objects	Back end-specific responses (1 per link)

We'd have to do a similar exercise for each such scenario identified before as required, but if we did, the impact on the factors would be:

The number of tests required to be run by humans has come down again and the tests run by machines can be run with lesser resources.
The cost of running each of the tests has reduced; so has the cost of maintaining test environments as live environments no longer need be maintained.
The resultant overall quality now depends largely on the fidelity with which the end-to-end scenarios have been mapped to component tests. If there is a high correlation between the two, overall quality should still remain within the original acceptable limits.

This last is not easy to do for various reasons:

There needs to be a coherent effort to ensure scenarios matchup between components; else component owners could define their sphere of influence differently from the scenarios deemed organizationally important.
The larger the size of mock data to be created, the more difficult it is to create it with high fidelity. Shortcuts such as record-replay mechanisms might help, but only if they've been sanitized to remove volatile data and then made generic to match the expected scenarios.
Ownership is shared; so division of labor should be clearly defined via policy. For example, each component owner should be responsible for her front end, ie, provide mock data and a set of regression test scripts to his clients; she can therefore rely on her dependents to do the same. Without such a policy, key pieces of the puzzle will be missing and the overall quality will suffer.

Optimizing testing - Round 4: Common Test Scenario Database

These issues with implementation of component testing may even lead to a regression back to manual testing. The crux of the problem is that the cohesive force of the end-to-end test is lost in component testing very easily.

The central idea with the common test scenario database is retain the benefits of component testing while bringing back that cohesion via data: we need to ensure that test data that is distributed across the various component test scripts still have the same tie-in to the original scenario. That way, every component owner in a particular scenario refers to the same scenario using the same language. While we're at it, it would also be beneficial to change the mock data in two ways:

Replace significant pieces of live data with values that stand for the class of test data that we will use in the specific scenario. E.g., the destination data item when used in a scenario where it represents a top destination could be given the canonical name "Top Dest1". Alternatively - assuming this is clearly understood company-wide - a real value can be used as a canonical one ; eg, in this case "Las Vegas" could stand for top destination; but then it shouldn't be used in any other scenario.
Clear out any recorded values from the mock data so only known values remain. This eliminates side effects from remnant data but requires a higher degree of discipline.

The larger change would be to introduce full-fledged exemplar data sets for application domain concepts that cannot be confused with live data, but clearly denote the exact scenario in which they can be used; and use policy to drive adoption of these exemplar data sets as the mock data backbone.

To continue on the hotel search example, the first step would be to define the following exemplar data:

Concept	Exemplar Data	Comment
Top Destination	LAS
Regular Destination	USDEST1
Special Destination	Acme Cozumel	Added "Acme" to Destination to call out that this is a test value
Next Week Friday	(Computed Value)	Mock data framework should be able to generate such values and respond appropriately
Hotel Chain	Acme Hotels LLC
Hotel	Grand Acme Ritz Chicago
Top Hotel @ Top Destination	Acme Hotel and Casino

The component test from before can then be rewritten like so:

Component	Given a test script that provides this input	..should provide this output	..using this mock data
Web-wl	LAS, Next fri, Next Sun, 1 guest	1st page has 25 hotels 1st hotel is Acme Hotel & Casino @ $100 2nd hotel is Acme MGM @ $125 …and so on	TBS response containing the 25 hotels (Note: other hosts required to bring wl up ignored for now)
TBS/Plugin	LAS, mm/dd/yyyy,mm/dd/yyyy,1 + other TBS required params	Arraylist of BookProduct objects	HSE response containing the 25 hotels
HSE	LAS, mm/dd/yyyy,mm/dd/yyyy,1 + other HSE required params	Arraylist of objects	Market/Markup response Supplier Link responses
SL Host(s)	LAS, mm/dd/yyyy,mm/dd/yyyy,1 + other Supplier Link required params	Arraylist of objects	SL-specific responses (1 per link)

More importantly, when a second scenario has to be mapped to component tests, the exemplar data table above should be checked to see if the concepts in that scenario already exist, and if so they should be reused.

So, to convert the following scenario:

Input	Expected Output
Top Destination , Peak Weekend, 4 guest	1st page has 25 hotels 1st hotel is Acme Hotel & Casino @ $100 2nd hotel is Acme MGM @ $125 …and so on

...into component tests, the following data items will have to be reused:

Top Destination (LAS)
Top Hotels (Acme Hotel & Casino, Acme MGM)

…and some new data items will have to be added:

Peak Weekend (Thanksgiving dates, for eg)

…which will further be reused when automating the scenario:

Input	Expected Output
Top Packaging Destination , Peak Weekend, n guests Acme Mexico Dest1 Labor Day Weekend dates 2 guests	1st page has 25 hotels 1st hotel is Acme Cozumel Resort1 @ $100 2nd hotel is Acme Cozumel Resort2 @ $125 …and so on

.. Which will further require new data items to be created, and so on.

When a new feature is added, say separate pricing for children or prepaid hotel rooms, that's the time for a completely new set of hotel chains and hotels to be created.

Over time, this practice of reusing test scenarios results in the creation of the Test Scenario Database which becomes the lingua franca across the organization when talking about Quality issues.

Let's see how our measuring factors score with this optimization:

The number of tests required to be run by humans hasn't changed since the last round.
The cost of running each of the tests remains as before. If there were any collateral increase in costs of using live environments due to inability to trust component tests, that is removed however.
The resultant overall quality still depends largely on the fidelity with which the end-to-end scenarios have been mapped to component tests; but there's a direct correlation possible because the organization now has the "common language" of the scenarios enshrined in the test data. This is the killer quality of the common scenario database.

Notes on implementation

Just as with automation, creating a Test Scenario Database is easier said than done. Policy from "up top" will certainly help; but a grassroots approach is also possible because the database can be built one scenario at a time. It does require a few enthusiastic converts, but some key component owners being convinced will create the kernel of good, reusable (and reused) test scenarios which can then be supported via policy.Once their use is common, network effects will take care of their popularity.
The Quality Organization should own the Test Scenario database and gate-keep use and reuse. Developers may create test scenarios, but should get approval from Q.
Component owners should be responsible for their component and their front end interface; and expect their dependents to do the same. That way they have the responsibility towards their clients and expect the same from their dependent servers.

Todo

Create a Test Scenario database tie-in for the Robot Framework