Testing Topic Guide

Having good tests in place is absolutely critical for ensuring a stable, maintainable codebase. Hopefully that doesn’t need any more explanation.

However, what defines a “good” test is not always obvious, and there are a lot of common pitfalls that can easily shoot your test suite in the foot.

If you already know everything about testing but are fed up with trying to debug why a specific test failed, you can skip the intro and jump straight to Debugging Unit Tests.

An overview of testing

There are three main types of tests, each with their associated pros and cons:

Unit tests

These are isolated, stand-alone tests with no external dependencies. They are written from the perspective of “knowing the code”, and test the assumptions of the codebase and the developer.

Pros:

  • Generally lightweight and fast.
  • Can be run anywhere, anytime since they have no external dependencies.

Cons:

  • Easy to be lax in writing them, or lazy in constructing them.
  • Can’t test interactions with live external services.

Functional tests

These are generally also isolated tests, though sometimes they may interact with other services running locally. The key difference between functional tests and unit tests, however, is that functional tests are written from the perspective of the user (who knows nothing about the code) and only knows what they put in and what they get back. Essentially this is a higher-level testing of “does the result match the spec?”.

Pros:

  • Ensures that your code always meets the stated functional requirements.
  • Verifies things from an “end user” perspective, which helps to ensure a high-quality experience.
  • Designing your code with a functional testing perspective in mind helps keep a higher-level viewpoint in mind.

Cons:

  • Requires an additional layer of thinking to define functional requirements in terms of inputs and outputs.
  • Often requires writing a separate set of tests and/or using a different testing framework from your unit tests.
  • Doesn’t offer any insight into the quality or status of the underlying code, only verifies that it works or it doesn’t.

Integration Tests

This layer of testing involves testing all of the components that your codebase interacts with or relies on in conjunction. This is equivalent to “live” testing, but in a repeatable manner.

Pros:

  • Catches many bugs that unit and functional tests will not.
  • Doesn’t rely on assumptions about the inputs and outputs.
  • Will warn you when changes in external components break your code.
  • Will take screenshot of the current page on test fail for easy debug

Cons:

  • Difficult and time-consuming to create a repeatable test environment.
  • Did I mention that setting it up is a pain?
Screenshot directory could be set through horizon.conf file, default value:
”./integration_tests_screenshots”

So what should I write?

A few simple guidelines:

  1. Every bug fix should have a regression test. Period.
  2. When writing a new feature, think about writing unit tests to verify the behavior step-by-step as you write the feature. Every time you’d go to run your code by hand and verify it manually, think “could I write a test to do this instead?”. That way when the feature is done and you’re ready to commit it you’ve already got a whole set of tests that are more thorough than anything you’d write after the fact.
  3. Write tests that hit every view in your application. Even if they don’t assert a single thing about the code, it tells you that your users aren’t getting fatal errors just by interacting with your code.

What makes a good unit test?

Limiting our focus just to unit tests, there are a number of things you can do to make your unit tests as useful, maintainable, and unburdensome as possible.

Test data

Use a single, consistent set of test data. Grow it over time, but do everything you can not to fragment it. It quickly becomes unmaintainable and perniciously out-of-sync with reality.

Make your test data as accurate to reality as possible. Supply all the attributes of an object, provide objects in all the various states you may want to test.

If you do the first suggestion above first it makes the second one far less painful. Write once, use everywhere.

To make your life even easier, if your codebase doesn’t have a built-in ORM-like function to manage your test data you can consider building (or borrowing) one yourself. Being able to do simple retrieval queries on your test data is incredibly valuable.

Mocking

Mocking is the practice of providing stand-ins for objects or pieces of code you don’t need to test. While convenient, they should be used with extreme caution.

Why? Because overuse of mocks can rapidly land you in a situation where you’re not testing any real code. All you’ve done is verified that your mocking framework returns what you tell it to. This problem can be very tricky to recognize, since you may be mocking things in setUp methods, other modules, etc.

A good rule of thumb is to mock as close to the source as possible. If you have a function call that calls an external API in a view , mock out the external API, not the whole function. If you mock the whole function you’ve suddenly lost test coverage for an entire chunk of code inside your codebase. Cut the ties cleanly right where your system ends and the external world begins.

Similarly, don’t mock return values when you could construct a real return value of the correct type with the correct attributes. You’re just adding another point of potential failure by exercising your mocking framework instead of real code. Following the suggestions for testing above will make this a lot less burdensome.

Assertions and verification

Think long and hard about what you really want to verify in your unit test. In particular, think about what custom logic your code executes.

A common pitfall is to take a known test object, pass it through your code, and then verify the properties of that object on the output. This is all well and good, except if you’re verifying properties that were untouched by your code. What you want to check are the pieces that were changed, added, or removed. Don’t check the object’s id attribute unless you have reason to suspect it’s not the object you started with. But if you added a new attribute to it, be damn sure you verify that came out right.

It’s also very common to avoid testing things you really care about because it’s more difficult. Verifying that the proper messages were displayed to the user after an action, testing for form errors, making sure exception handling is tested... these types of things aren’t always easy, but they’re extremely necessary.

To that end, Horizon includes several custom assertions to make these tasks easier. assertNoFormErrors(), assertMessageCount(), and assertNoMessages() all exist for exactly these purposes. Moreover, they provide useful output when things go wrong so you’re not left scratching your head wondering why your view test didn’t redirect as expected when you posted a form.

Debugging Unit Tests

Tips and tricks

  1. Use assertNoFormErrors() immediately after your client.post call for tests that handle form views. This will immediately fail if your form POST failed due to a validation error and tell you what the error was.

  2. Use assertMessageCount() and assertNoMessages() when a piece of code is failing inexplicably. Since the core error handlers attach user-facing error messages (and since the core logging is silenced during test runs) these methods give you the dual benefit of verifying the output you expect while clearly showing you the problematic error message if they fail.

  3. Use Python’s pdb module liberally. Many people don’t realize it works just as well in a test case as it does in a live view. Simply inserting import pdb; pdb.set_trace() anywhere in your codebase will drop the interpreter into an interactive shell so you can explore your test environment and see which of your assumptions about the code isn’t, in fact, flawlessly correct.

  4. If the error is in the Selenium test suite, you’re likely getting very little information about the error. To increase the information provided to you, edit horizon/test/settings.py to set DEBUG = True and set the logging level to ‘DEBUG’ for the default ‘test’ logger. Also, add a logger config for Django:

         },
         'loggers': {
    +        'django': {
    +            'handlers': ['test'],
    +            'propagate': False,
    +        },
             'django.db.backends': {
    

Common pitfalls

There are a number of typical (and non-obvious) ways to break the unit tests. Some common things to look for:

  1. Make sure you stub out the method exactly as it’s called in the code being tested. For example, if your real code calls api.keystone.tenant_get, stubbing out api.tenant_get (available for legacy reasons) will fail.
  2. When defining the expected input to a stubbed call, make sure the arguments are identical, this includes str vs. int differences.
  3. Make sure your test data are completely in line with the expected inputs. Again, str vs. int or missing properties on test objects will kill your tests.
  4. Make sure there’s nothing amiss in your templates (particularly the {% url %} tag and its arguments). This often comes up when refactoring views or renaming context variables. It can easily result in errors that you might not stumble across while clicking around the development server.
  5. Make sure you’re not redirecting to views that no longer exist, e.g. the index view for a panel that got combined (such as instances & volumes).
  6. Make sure your mock calls are in order before calling mox.ReplayAll. The order matters.
  7. Make sure you repeat any stubbed out method calls that happen more than once. They don’t automatically repeat, you have to explicitly define them. While this is a nuisance, it makes you acutely aware of how many API calls are involved in a particular function.

Understanding the output from mox

Horizon uses mox as its mocking framework of choice, and while it offers many nice features, its output when a test fails can be quite mysterious.

Unexpected Method Call

This occurs when you stubbed out a piece of code, and it was subsequently called in a way that you didn’t specify it would be. There are two reasons this tends to come up:

  1. You defined the expected call, but a subtle difference crept in. This may be a string versus integer difference, a string versus unicode difference, a slightly off date/time, or passing a name instead of an id.
  2. The method is actually being called multiple times. Since mox uses a call stack internally, it simply pops off the expected method calls to verify them. That means once a call is used once, it’s gone. An easy way to see if this is the case is simply to copy and paste your method call a second time to see if the error changes. If it does, that means your method is being called more times than you think it is.

Expected Method Never Called

This one is the opposite of the unexpected method call. This one means you told mox to expect a call and it didn’t happen. This is almost always the result of an error in the conditions of the test. Using the assertNoFormErrors() and assertMessageCount() will make it readily apparent what the problem is in the majority of cases. If not, then use pdb and start interrupting the code flow to see where things are getting off track.

Integration tests in Horizon

The integration tests currently live in the Horizon repository, see here, which also contains instructions on how to run the tests. To make integration tests more understandable and maintainable, the Page Object pattern is used throughout them.

Warning

To enable integration tests support before running them, please copy openstack_dashboard/local/local_settings.d/_20_integration_tests_scaffolds.py.example to openstack_dashboard/local/local_settings.d/_20_integration_tests_scaffolds.py and then run ./manage.py collectstatic –clear && ./manage.py compress.

Horizon repository also provides two shell scripts, which are executed in pre_test_hook and post_test_hook respectively. Pre hook is generally used for modifying test environment, while post hook is used for running actual integration tests with tox and collecting test artifacts. Thanks to the incorporating all modifications to tests into Horizon repository, one can alter both tests and test environment and see the immediate results in Jenkins job output.

Page Object pattern

Within any web application’s user interface (UI) there are areas that the tests interact with. A Page Object simply models these as objects within the test code. This reduces the amount of duplicated code; if the UI changes, the fix needs only be applied in one place.

Page Objects can be thought of as facing in two directions simultaneously. Facing towards the developer of a test, they represent the services offered by a particular page. Facing away from the developer, they should be the only thing that has a deep knowledge of the structure of the HTML of a page (or part of a page). It is simplest to think of the methods on a Page Object as offering the “services” that a page offers rather than exposing the details and mechanics of the page. As an example, think of the inbox of any web-based email system. Amongst the services that it offers are typically the ability to compose a new email, to choose to read a single email, and to list the subject lines of the emails in the inbox. How these are implemented should not matter to the test.

Writing reusable and maintainable Page Objects

Because the main idea is to encourage the developer of a test to try and think about the services that they are interacting with rather than the implementation, Page Objects should seldom expose the underlying WebDriver instance. To facilitate this, methods on the Page Object should return other Page Objects. This means that we can effectively model the user’s journey through the application.

Another important thing to mention is that a Page Object need not represent an entire page. It may represent a section that appears many times within a site or page, such as site navigation. The essential principle is that there is only one place in your test suite with knowledge of the structure of the HTML of a particular (part of a) page. With this in mind, a test developer builds up regions that become reusable components (example of a base form). These properties can then be redefined or overridden (e.g. selectors) in the actual pages (subclasses) (example of a tabbed form).

The page objects are read-only and define the read-only and clickable elements of a page, which work to shield the tests. For instance, from the test perspective, if “Logout” used to be a link but suddenly becomes an option in a drop-down menu, there are no changes (in the test itself) because it still simply calls the “click_on_logout” action method.

This approach has two main aspects:

  • The classes with the actual tests should be as readable as possible
  • The other parts of the testing framework should be as much about data as possible, so that if the CSS etc. changes you only need to change that one property. If the flow changes, only the action method should need to change.

There is little that is Selenium-specific in the Pages, except for the properties. There is little coupling between the tests and the pages. Writing the tests becomes like writing out a list of steps (by using the previously mentioned action methods). One of the key points, particularly important for this kind of UI driven testing is to isolate the tests from what is behind them.

Debugging integration tests

Even perfectly designed Page Objects are not a guarantee that your integration test will not ever fail. This can happen due to different causes:

The first and most anticipated kind of failure is the inability to perform a testing scenario by a living person simply because some OpenStack service or Horizon itself prevents them from doing so. This is exactly the kind that integration tests are designed to catch. Let us call them “good” failures.

All other kinds of failures are unwanted and could be roughly split into the two following categories:

  1. The failures that occur due to changes in application’s DOM. some CSS/ Xpath selectors no longer matching Horizon app’s DOM. The usual signature for that kind of failures is having a DOM changing patch for which the test job fails with a message like this selenium.common.exceptions.NoSuchElementException: Message: Unable to locate element: {“method”:”css selector”,”selector”:”div.modal-dialog”}. If you find yourself in such a situation, you should fix the Page Object selectors according to the DOM changes you made.
  2. Unfortunately it is still quite possible to get the above error for a patch which didn’t implement any DOM changes. Among the reasons of such behavior observed in past were:
    • Integration tests relying on relative ordering of form fields and table actions that broke with the addition of a new field. This issue should be fixed by now, but may reappear in future for different entities.
    • Integration tests relying on popups disappearing by the time a specific action needs to be taken (or not existing at all). This expectation turned out to be very fragile, since the speed of tests execution by Jenkins workers may change independently of integration test code (hence, popups disappear too late to free the way for the next action). The unexpected (both too long and too short) timeouts aren’t limited to just popups, but apply to every situation when the element state transition is not instant (like opening an external link, going to another page in Horizon, waiting for button to become active, waiting for a table row to change its state). Luckily, most transitions of “element becomes visible/ emerge to existence from non-existence” kind are already bulletproofed using implicit_wait parameter in integration_tests/horizon.conf file. Selenium just waits for specified amount of seconds for an element to become visible (if it’s not already visible) giving up when it exceeds (with the above error). Also it’s worth mentioning explicit_wait parameter which is considered when the selenium wait_until method is involved (and it is used, e.g. in waiting for spinner and messages popups to disappear).

An inconvenient thing about reading test results in the console.html file attached to every gate-horizon-dsvm-integration finished job is that the test failure may appear either as failure (assertion failed), or as error (expected element didn’t show up). In both cases an inquirer should suspect a legitimate failure first (i.e., treat errors as failures). Unfortunately, no clear method exists for the separation of “good” from “bad” failures. Each case is unique and full of mysteries.

The Horizon testing mechanism tries to alleviate this ambiguity by providing several facilities to aid in failure investigation:

  • First there comes a screenshot made for every failed test (in a separate folder, on a same level as console.html) - almost instant snapshot of a screen on the moment of failure (almost sometimes matters, especially in a case of popups that hang on a screen for a limited time);
  • Then the patient inquirer may skim through the vast innards of console.html, looking at browser log first (all javascript and css errors should come there),
  • Then looking at a full textual snapshot of a page for which test failed (sometimes it gives a more precise picture than a screenshot),
  • And finally looking at test error stacktrace (most useful) and a lengthy output of requests/ responses with a selenium server. The last log sometimes might tell us how long a specific web element was polled before failing (in case of implicit_wait there should be a series of requests to the same element).

The best way to solve the cause of test failure is running and debugging the troublesome test locally. You could use pdb or Python IDE of your choice to stop test execution in arbitrary points and examining various Page Objects attributes to understand what they missed. Looking at the real page structure in browser developer tools also could explain why the test fails. Sometimes it may be worth to place breakpoints in JavaScript code (provided that static is served uncompressed) to examine the objects of interest. If it takes long, you may also want to increase the webdriver’s timeout so it will not close browser windows forcefully. Finally, sometimes it may make sense to examine the contents of logs directory, especially apache logs - but that is mostly the case for the “good” failures.

Writing your first integration test

So, you are going to write your first integration test and looking for some guidelines on how to do it. The first and the most comprehensive source of knowledge is the existing codebase of integration tests. Look how other tests are written, which Page Objects they use and learn by copying. Accurate imitation will eventually lead to a solid understanding. Yet there are few things that may save you some time when you know them in advance.

File and directory layout and go_to_*page() methods

Below is the filesystem structure that test helpers rely on.:

horizon/
└─ openstack_dashboard/
   └─ test/
      └─ integration_tests/
         ├─ pages/
         │  ├─ admin/
         │  │  ├─ __init__.py
         │  │  └─ system/
         │  │     ├─ __init__.py
         │  │     └─ flavorspage.py
         │  ├─ project/
         │  │  └─ compute/
         │  │     ├─ __init__.py
         │  │     ├─ access_and_security/
         │  │     │  ├─ __init__.py
         │  │     │  └─ keypairspage.py
         │  │     └─ imagespage.py
         │  └─ navigation.py
         ├─ regions/
         ├─ tests/
         ├─ config.py
         └─ horizon.conf

New tests are put into integration_tests/tests, where they are grouped by the kind of entities being tested (test_instances.py, test_networks.py, etc). All Page Objects to be used by tests are inside pages/directory, the nested directory structure you see within it obeys the value of Navigation.CORE_PAGE_STRUCTURE you can find at pages/navigation.py module. The contents of the CORE_PAGE_STRUCTURE variable should in turn mirror the structure of standard dashboard sidebar menu. If this condition is not met, the go_to_<pagename>page() methods which are generated automatically at runtime will have problems matching the real sidebar items. How are these go_to_*page() methods are generated? From the sidebar’s point of view, dashboard content could be at most four levels deep: Dashboard, Panel Group, Panel and Tab. Given the mixture of these entities in existing dashboards, it was decided that:

  • When panels need to be addressed with go_to_<pagename>page() methods, two components in the method’s name are enough for distinguishing the right path to go along, namely a Panel name and a Panel Group name (or a Dashboard name, if no Panel Group exists above Panel). For example,
    • go_to_system_flavorspage() method to go to Admin->System->Flavors and
    • go_to_identity_projectspage() method to go to Identity->Projects panel.
  • When we need to go one level deeper, i.e. go to the specific TableTab on any panel that has several tabs, three components are enough - Panel Group, Panel and Tab names. For example, go_to_compute_accessandsecurity_floatingipspage() for navigating to Project->Compute->Access & Security->Floating IPs tab. Note that one cannot navigate to a Panel level if that Panel has several tabs (i.e., only terminal levels could be navigated to).

As you might have noticed, method name components are chosen from normalized items of the CORE_PAGE_STRUCTURE dictionary, where normalization means replacing spaces with _ symbol and & symbol with and, then downcasing all symbols.

Once the go_to_*page() method’s name is parsed and the proper menu item is matched in a dashboard, it should return the proper Page Object. For that to happen a properly named class should reside in a properly named module located in the right place of the filesystem. More specifically and top down:

  1. Page Object class is located in:
    • <dashboard>/<panel_group>/<panel>page.py file for non-tabbed pages
    • <dashboard>/<panel_group>/<panel>/<tab>page.py file for tabbed pages Values <dashboard>, <panel_group>, <panel> and <tab> are the normalized versions of the items from the CORE_PAGE_STRUCTURE dictionary.
  2. Within the above module a descendant of basepage.BaseNavigationPage should be defined, its name should have the form <Panel>Page or <Tab>Page, where <Panel> and <Tab> are capitalized versions of normalized <panel> and <tab> items respectively.

Reusable regions

  • TableRegion binds to the HTML Horizon table using the TableRegion‘s name attribute. To bind to the proper table this attribute has to be the same as the name attribute of a Meta subclass of a corresponding tables.DataTable descendant in the Python code. TableRegion provides all the needed facilities for solving the following table-related tasks.

    • Getting a specific row from a table matched by the column name and a target text within that column (use get_row() method) or taking all the existing rows on a current table page with rows property.
    • Once you have a reference to a specific row, it can either be marked with mark() for further batch actions or split to cells (using cells property which is dictionary representing column name as a key to cell wrapper as a value).
    • For interacting with actions TableRegion provides 2 decorators, namely @bind_table_action() and @bind_row_action() which bind to the actual HTML button widget and decorate the specific table methods. These methods in turn should click a bound button (comes as these methods’ second argument after self) and usually return a new region which is most often bound to a modal form being shown after clicking that button in real Horizon.
    • Another important part of TableRegion are the facilities for checking the properties of a paged table - assert_definition(), is_next_link_available() and is_prev_link_available() helpers and turn_next_page() / turn_prev_page() which obviously cause the next / prev table page to be shown.
  • when interacting with modal and non-modal forms three flavors of form wrappers can be used.

    • BaseFormRegion is used for simplest forms which are usually ‘Submit’ / ‘Cancel’ dialogs with no fields to be filled.

    • FormRegion is the most used wrapper which provides interaction with the fields within that form. Every field is backed by its own wrapper class, while the FormRegion acts as a container which initializes all the field wrappers in its __init__() method. Field mappings passed to __init__() could be

      • either a tuple of string labels, in that case the same label is used for referencing the field in test code and for binding to the HTML input (should be the same as name attribute of that widget, could be seen in Django code defining that form in Horizon)
      • or a dictionary, where the key will be used for referencing the test field and the value will be used for binding to the HTML input. Also it is feasible to provide values other than strings in that dictionary - in this case they are meant to be a Python class. This Python class will be initialized as any BaseRegion is usually initialized and then the value’s key will be used for referencing this object. This is useful when dealing with non-standard widgets in forms (like Membership widget in Create/​Edit Project form or Networks widget in Launch Instance form).
    • TabbedFormRegion is a slight variation of FormRegion, it has several tabs and thus can accept a tuple of tuples / dictionaries of field mappings, where every tuple corresponds to a tab of a real form, binding order is that first tuple binds to leftmost tab, which has index 0. Passing default_tab other than 0 to TabbedFormRegion.__init__ we can make the test form to be created with the tab other than the leftmost being shown immediately. Finally the method switch_to allows us to switch to any existing form’s tab.

  • MessageRegion is a small region, but is very important for asserting that everything goes well in Horizon under test. Technically, the find_message_and_dismiss method belongs to BasePage class, but whenever it is called, regions.messages module is imported as well to pass a messages.SUCCESS / messages.ERROR argument into. The method returns True / False depending on if the specified message was found and dismissed (which could be then asserted for).

Customizing tests to a specific gate environment

  • Upstream gate environment is not the only possible environment where Horizon integration tests can be run. Various downstream distributions may also want to run them. To ease the adoption of upstream tests to possibly different conditions of a downstream gate environment, integration tests use a configuration machinery backed by oslo.config library. It includes the following pieces of knowledge:
    • integration_tests/config.py file where all possible setting groups and settings are defined along with their descriptions and defaults. If you are going to add a new setting to Horizon integration tests, you should add it first to this file.
    • integration_tests/horizon.conf file - where all the overrides are actually located. For clarity its contents mirrors the default values in config.py (although technically they could be completely commented out with the same result).
    • To make developers’ lives easier a local-only (not tracked by git) counterpart of horizon.conf could exist at the same directory, named ‘local-horizon.conf’. It is meant solely for overriding values from horizon.conf that a developer’s environment might differ from the gate environment (like Horizon url or admin user password).
  • When integration tests are run by openstack-infra/devstack-gate scripts they use 2 hooks to alter the devstack gate environment, namely pre_test_hook and post_test_hook. Contents of both hooks are defined inside the corresponding shell scripts located at ‘tools/gate/integration’ at the top-level of horizon repo. If you find yourself asking which of the hooks you need to modify - pre or post, keep the following things in mind.
    • Pre hook is executed before the Devstack is deployed, that essentially means that almost none of packages that are installed as OpenStack services dependencies during Devstack deployment are going to be present in the system. Yet all the repositories contained with PROJECTS variable defined in devstack-vm-gate-wrap.sh script will be already cloned by the moment pre hook is executed. So the natural use for it is to customize some Horizon settings before they are used in operations like compressing statics etc. That is how it is actually used now: it sets settings variable INTEGRATION_TESTS_SUPPORT to True, so all the support code for integration tests is included into Horizon static assets. If this variable was set to False, we would not be able to run integration tests.
    • Post hook is executed after Devstack is deployed, so integration tests themselves are run inside that hook, as well as various test artifacts collection. When you modify it, do not forget to save the exit code of a tox integration tests run and emit at the end of the script - or you may lose the SUCCESS/FAILURE status of the whole tests suite and tamper with the job results!

Writing integration tests for Horizon plugins

First, for more details on writing a Horizon plugin please refer to Horizon Plugin. Second, there are 2 possible setups when running integration tests for Horizon plugins.

The first setup, which is suggested to be used in gate of *-dashboard plugins is to get horizon as a dependency of a plugin and then run integration tests using horizon.conf config file inside the plugin repo. This way the plugin augments the location of Horizon built-in Page Objects with the location of its own Page Objects, contained within the plugin_page_path option and the Horizon built-in nav structure with its own nav structure contained within plugin_page_structure. Then the plugin integration tests are run against core Horizon augmented with just this particular plugin content.

The second setup may be used when it is needed to run integration tests for Horizon + several plugins. In other words, content from several plugins is merged into core Horizon content, then the combined integration tests from core Horizon and all the involved plugins are run against the resulting dashboards. To make this possible both options plugin_page_path and plugin_page_structure have MultiStrOpt type. This means that they may be defined several times and all the specified values will be gathered in a list, which is iterated over when running integration tests. In this setup it’s easier to run the tests from Horizon repo, using the horizon.conf file within it.

Also keep in mind that plugin_page_structure needs to be a strict JSON string, w/o trailing commas etc.