Developer’s Manual

Helping Ghini development

Installing Ghini always includes downloading the sources, connected to the github repository. This is so because in our eyes, every user is always potentially also a developer.

If you want to contribute to Ghini, you can do so in quite a few different ways:

  • Use the software, note the things you don’t like, open an issue for each of them. A developer will react sooner than you can imagine.
  • If you have an idea of what you miss in the software but can’t quite formalize it into separate issues, you could consider hiring a professional. This is the best way to make sure that something happens quickly on Ghini. Do make sure the developer opens issues and publishes their contribution on github.
  • Translate! Any help with translations will be welcome, so please do! you can do this without installing anything on your computer, just using the on-line translation service offered by
  • fork the respository, choose an issue, solve it, open a pull request. See the bug solving workflow below.

If you haven’t yet installed Ghini, and want to have a look at its code history, you can open our github project page and see all that has been going on around Ghini since its inception as Bauble, back in the year 2004.

Software source, versions, branches

If you want a particular version of Ghini, we release and maintain versions as branches. You should git checkout the branch corresponding to the version of your choice.

production line

Branch names for Ghini stable (production) versions are of the form ghini-x.y (eg: ghini-1.0); branch names where Ghini testing versions are published are of the form ghini-x.y-dev (eg: ghini-1.0-dev).

Development Workflow

Our workflow is to continuously commit to the testing branch, to often push them to github, to let travis-ci and check the quality of the pushed testing branches, finally, from time to time, to merge the testing branch into the corresponding release.

When working at larger issues, which seem to take longer than a couple of days, I might open a branch associated to the issue. I don’t do this very often.

larger issues

When facing a single larger issue, create a branch tag at the tip of a main development line (e.g.: ghini-1.0-dev), and follow the workflow described at

in short:

git up
git checkout -b issue-xxxx
git push origin issue-xxxx

Work on the new temporary branch. When ready, go to github, merge the branch with the main development line from which you branched, solve conflicts where necessary, delete the temporary branch.

When ready for publication, merge the development line into the corresponding production line.

Updating the set of translatable strings

From time to time, during the process of updating the software, you will be adding or modifying strings in the python sources, in the documentation, in the glade sources. Most of our strings are translatable, and are offered to weblate for people to contribute, in the form of several .po files.

A po is mostly composed of pairs of text portions, original and translation, and is specific to a target language. When a translator adds a translation on weblate, this reaches our repository on github. When a programmer adds a string to the software, this reaches weblate as “to be translated”.

Weblate hosts the Ghini project. Within this project we have components, each of which corresponds to a branch of a repository on github. Each component accepts translations in several languages.

component repository branch
Desktop 1.0 ghini.desktop ghini-1.0-dev
Desktop 1.1 ghini.desktop ghini-1.1-dev
Documentation 1.0 ghini.desktop-docs.i18n ghini-1.0-dev
Documentation 1.1 ghini.desktop-docs.i18n ghini-1.1-dev
Web 1.2 ghini.web master

To update the po files relative to the software, you set the working directory to the root of your checkout of ghini.desktop, and you run the script:


To update the po files relative to the documentation, you set the working directory to the root of your checkout of ghini.desktop-docs.i18n, and you run the script:


When you run either of the above mentioned scripts, chances are you need to commit all po files in the project. You may want to review the changes before committing them to the respository. This is most important when you perform a marginal correction to a string, like removing a typo.

Our documentation on readthedocs has an original English version, and several translations. We just follow the description of localisation, there’s nothing that we invented ourselves here.

Readthedocs checks the project’s Language setting, and invokes sphinx-intl to produce the formatted documentation in the target language. With the default configuration —which we did not alter— sphinx-intl expects one po file per source document, named as the source document, and that they all reside in the directory local/$(LANG)/LC_MESSAGES/.

On the other hand, Weblate (and ourselves) prefers a single po file per language, and keeps them all in the same /po directory, just as we do for the software project: /po/$(LANG).po.

In order not to repeat information, and to let both systems work their natural way, we have two sets of symbolic links (git honors them).

To summarise: when a file in the documentation is updated, the script will:

  1. copy the rst files from the software to the documentation;
  2. create a new pot file for each of the documentation files;
  3. merge all pot files into one doc.pot;
  4. use the updated doc.pot to update all doc.po files (one per language);
  5. create all symbolic links:
    1. those expected by sphinx-intl in /local/$(LANG)/LC_MESSAGES/
    2. those used by weblate in /po/$(LANG).po

We could definitely write the above in a Makefile, or even better include it in /doc/Makefile. Who knows, maybe we will do that.

Adding missing unit tests

If you are interested contributing to development of Ghini, a good way to do so would be by helping us finding and writing the missing unit tests.

A well tested function is one whose behaviour you cannot change without breaking at least one unit test.

We all agree that in theory theory and practice match perfectly and that one first writes the tests, then implements the function. In practice, however, practice does not match theory and we have been writing tests after writing and even publishing the functions.

This section describes the process of adding unit tests for

What to test

First of all, open the coverage report index, and choose a file with low coverage.

For this example, run in October 2015, we landed on, at 33%.

The first two functions which need tests, edit_callback and add_genera_callback, include creation and activation of an object relying on a custom dialog box. We should really first write unit tests for that class, then come back here.

The next function, remove_callback, also activates a couple of dialog and message boxes, but in the form of invoking a function requesting user input via yes-no-ok boxes. These functions we can easily replace with a function mocking the behaviour.

how to test

So, having decided what to describe in unit test, we look at the code and we see it needs discriminate a couple of cases:

parameter correctness
  • the list of families has no elements.
  • the list of families has more than one element.
  • the list of families has exactly one element.
  • the family has no genera
  • the family has one or more genera
  • the user confirms deletion
  • the user does not confirm deletion
  • all goes well when deleting the family
  • there is some error while deleting the family

I decide I will only focus on the cascade and the confirm aspects. Two binary questions: 4 cases.

where to put the tests

Locate the test script and choose the class where to put the extra unit tests.

what about skipped tests

The FamilyTests class contains a skipped test, implementing it will be quite a bit of work because we need rewrite the FamilyEditorPresenter, separate it from the FamilyEditorView and reconsider what to do with the FamilyEditor class, which I think should be removed and replaced with a single function.

writing the tests

After the last test in the FamilyTests class, I add the four cases I want to describe, and I make sure they fail, and since I’m lazy, I write the most compact code I know for generating an error:

def test_remove_callback_no_genera_no_confirm(self):

def test_remove_callback_no_genera_confirm(self):

def test_remove_callback_with_genera_no_confirm(self):

def test_remove_callback_with_genera_confirm(self):

One test, step by step

Let’s start with the first test case.

When writing tests, I generally follow the pattern:

  • T₀ (initial condition),
  • action,
  • T₁ (testing the result of the action given the initial conditions)

what’s in a name — unit tests

There’s a reason why unit tests are called unit tests. Please never test two actions in one test.

So let’s describe T₀ for the first test, a database holding a family without genera:

def test_remove_callback_no_genera_no_confirm(self):
    f5 = Family(family=u'Arecaceae')

We do not want the function being tested to invoke the interactive utils.yes_no_dialog function, we want remove_callback to invoke a non-interactive replacement function. We achieve this simply by making utils.yes_no_dialog point to a lambda expression which, like the original interactive function, accepts one parameter and returns a boolean. In this case: False:

def test_remove_callback_no_genera_no_confirm(self):
    # T_0
    f5 = Family(family=u'Arecaceae')

    # action
    utils.yes_no_dialog = lambda x: False
    from import remove_callback

Next we test the result.

Well, we don’t just want to test whether or not the object Arecaceae was deleted, we also should test the value returned by remove_callback, and whether yes_no_dialog and message_details_dialog were invoked or not.

A lambda expression is not enough for this. We do something apparently more complex, which will make life a lot easier.

Let’s first define a rather generic function:

def mockfunc(msg=None, name=None, caller=None, result=None):
    caller.invoked.append((name, msg))
    return result

and we grab partial from the functools standard module, to partially apply the above mockfunc, leaving only msg unspecified, and use this partial application, which is a function accepting one parameter and returning a value, to replace the two functions in utils. The test function now looks like this:

def test_remove_callback_no_genera_no_confirm(self):
    # T_0
    f5 = Family(family=u'Arecaceae')
    self.invoked = []

    # action
    utils.yes_no_dialog = partial(
        mockfunc, name='yes_no_dialog', caller=self, result=False)
    utils.message_details_dialog = partial(
        mockfunc, name='message_details_dialog', caller=self)
    from import remove_callback
    result = remove_callback([f5])

The test section checks that message_details_dialog was not invoked, that yes_no_dialog was invoked, with the correct message parameter, that Arecaceae is still there:

# effect
self.assertFalse('message_details_dialog' in
                 [f for (f, m) in self.invoked])
self.assertTrue(('yes_no_dialog', u'Are you sure you want to '
                 'remove the family <i>Arecaceae</i>?')
                in self.invoked)
self.assertEquals(result, None)
q = self.session.query(Family).filter_by(family=u"Arecaceae")
matching = q.all()
self.assertEquals(matching, [f5])

And so on

there are two kinds of people, those who complete what they start, and so on

Next test is almost the same, with the difference that the utils.yes_no_dialog should return True (this we achieve by specifying result=True in the partial application of the generic mockfunc).

With this action, the value returned by remove_callback should be True, and there should be no Arecaceae Family in the database any more:

def test_remove_callback_no_genera_confirm(self):
    # T_0
    f5 = Family(family=u'Arecaceae')
    self.invoked = []

    # action
    utils.yes_no_dialog = partial(
        mockfunc, name='yes_no_dialog', caller=self, result=True)
    utils.message_details_dialog = partial(
        mockfunc, name='message_details_dialog', caller=self)
    from import remove_callback
    result = remove_callback([f5])

    # effect
    self.assertFalse('message_details_dialog' in
                     [f for (f, m) in self.invoked])
    self.assertTrue(('yes_no_dialog', u'Are you sure you want to '
                     'remove the family <i>Arecaceae</i>?')
                    in self.invoked)
    self.assertEquals(result, True)
    q = self.session.query(Family).filter_by(family=u"Arecaceae")
    matching = q.all()
    self.assertEquals(matching, [])

have a look at commit 734f5bb9feffc2f4bd22578fcee1802c8682ca83 for the other two test functions.

Testing logging

Our bauble.test.BaubleTestCase objects use handlers of the class bauble.test.MockLoggingHandler. Every time an individual unit test is started, the setUp method will create a new handler and associate it to the root logger. The tearDown method takes care of removing it.

You can check for presence of specific logging messages in self.handler.messages. messages is a dictionary, initially empty, with two levels of indexation. First the name of the logger issuing the logging record, then the name of the level of the logging record. Keys are created when needed. Values hold lists of messages, formatted according to whatever formatter you associate to the handler, defaulting to logging.Formatter("%(message)s").

You can explicitly empty the collected messages by invoking self.handler.clear().

Putting all together

From time to time you want to activate the test class you’re working at:

nosetests bauble/plugins/plants/

And at the end of the process you want to update the statistics:


Structure of user interface

The user interface is built according to the ModelViewPresenter architectural pattern. For much of the interface, Model is a SQLAlchemy database object, but we also have interface elements where there is no corresponding database model. In general:

  • The View is described as part of a glade file. This should include the signal-callback and ListStore-TreeView associations. Just reuse the base class GenericEditorView defined in bauble.editor. When you create your instance of this generic class, pass it the glade file name and the root widget name, then hand this instance over to the presenter constructor.

    In the glade file, in the action-widgets section closing your GtkDialog object description, make sure every action-widget element has a valid response value. Use valid GtkResponseType values, for example:


    There is no easy way to unit test a subclassed view, so please don’t subclass views, there’s really no need to.

    In the glade file, every input widget should define which handler is activated on which signal. The generic Presenter class offers generic callbacks which cover the most common cases.

    • GtkEntry (one-line text entry) will handle the changed signal, with either on_text_entry_changed or on_unique_text_entry_changed.
    • GtkTextView: associate it to a GtkTextBuffer. To handle the changed signal on the GtkTextBuffer, we have to define a handler which invokes the generic on_textbuffer_changed, the only role for this function is to pass our generic handler the name of the model attribute that receives the change. This is a workaroud for an unresolved bug in GTK.
    • GtkComboBox with translated texts can’t be easily handled from the glade file, so we don’t even try. Use the init_translatable_combo method of the generic GenericEditorView class, but please invoke it from the presenter.
  • The Model is just an object with known attributes. In this interaction, the model is just a passive data container, it does nothing more than to let the presenter modify it.

  • The subclassed Presenter defines and implements:

    • widget_to_field_map, a dictionary associating widget names to name of model attributes,
    • view_accept_buttons, the list of widget names which, if activated by the user, mean that the view should be closed,
    • all needed callbacks,
    • optionally, it plays the model role, too.

    The presenter continuously updates the model according to changes in the view. If the model corresponds to a database object, the presenter commits all model updates to the database when the view is closed successfully, or rolls them back if the view is canceled. (this behaviour is influenced by the parameter do_commit)

    If the model is something else, then the presenter will do something else.


    A well behaved presenter uses the view api to query the values inserted by the user or to forcibly set widget statuses. Please do not learn from the practice of our misbehaving presenters, some of which directly handle fields of view.widgets. By doing so, these presenters prevents us from writing unit tests.

The base class for the presenter, GenericEditorPresenter defined in bauble.editor, implements many useful generic callbacks. There is a MockView class, that you can use when writing tests for your presenters.


Contact and ContactPresenter are implemented following the above lines. The view is defined in the file.

A good example of Presenter/View pattern (no model) is given by the connection manager.

We use the same architectural pattern for non-database interaction, by setting the presenter also as model. We do this, for example, for the JSON export dialog box. The following command will give you a list of GenericEditorView instantiations:

grep -nHr -e GenericEditorView\( bauble

Extending Ghini with Plugins

Nearly everything about Ghini is extensible through plugins. Plugins can create tables, define custom searchs, add menu items, create custom commands and more.

To create a new plugin you must extend the bauble.pluginmgr.Plugin class.

The Tag plugin is a good minimal example, even if the TagItemGUI falls outside the Model-View-Presenter architectural pattern.

Plugins structure

Ghini is a framework for handling collections, and is distributed along with a set of plugins making Ghini a botanical collection manager. But Ghini stays a framework and you could in theory remove all plugins we distribute and write your own, or write your own plugins that extend or complete the current Ghini behaviour.

Once you have selected and opened a database connection, you land in the Search window. The Search window is an interaction between two objects: SearchPresenter (SP) and SearchView (SV).

SV is what you see, SP holds the program status and handles the requests you express through SV. Handling these requests affect the content of SV and the program status in SP.

The search results shown in the largest part of SV are rows, objects that are instances of classes registered in a plugin.

Each of these classes must implement an amount of functions in order to properly behave within the Ghini framework. The Ghini framework reserves space to pluggable classes.

SP knows of all registered (plugged in) classes, they are stored in a dictionary, associating a class to its plugin implementation. SV has a slot (a gtk.Box) where you can add elements. At any time, at most only one element in the slot is visible.

A plugin defines one or more plugin classes. A plugin class plays the role of a partial presenter (pP - plugin presenter) as it implement the callbacks needed by the associated partial view fitting in the slot (pV - plugin view), and the MVP pattern is completed by the parent presenter (SP), again acting as model. To summarize and complete:

  • SP acts as model,
  • the pV partial view is defined in a glade file.
  • the callbacks implemented by pP are referenced by the glade file.
  • a context menu for the SP row,
  • a children property.

when you register a plugin class, the SP:

  • adds the pV in the slot and makes it non-visible.
  • adds an instance of pP in the registered plugin classes.
  • tells the pP that the SP is the model.
  • connects all callbacks from pV to pP.

when an element in pV triggers an action in pP, the pP can forward the action to SP and can request SP that it updates the model and refreshes the view.

When the user selects a row in SP, SP hides everything in the pluggable slot and shows only the single pV relative to the type of the selected row, and asks the pP to refresh the pV with whatever is relative to the selected row.

Apart from setting the visibility of the various pV, nothing needs be disabled nor removed: an invisible pV cannot trigger events!

bug solving workflow

normal development workflow

  • while using the software, you notice a problem, or you get an idea of something that could be better, you think about it good enough in order to have a very clear idea of what it really is, that you noticed. you open an issue and describe the problem. someone might react with hints.
  • you open the issues site and choose one you want to tackle.
  • assign the issue to yourself, this way you are informing the world that you have the intention to work at it. someone might react with hints.
  • optionally fork the repository in your account and preferably create a branch, clearly associated to the issue.
  • write unit tests and commit them to your branch (please do not push failing unit tests to github, run nosetests locally first).
  • write more unit tests (ideally, the tests form the complete description of the feature you are adding or correcting).
  • make sure the feature you are adding or correcting is really completely described by the unit tests you wrote.
  • make sure your unit tests are atomic, that is, that you test variations on changes along one single variable. do not give complex input to unit tests or tests that do not fit on one screen (25 lines of code).
  • write the code that makes your tests succeed.
  • update the i18n files (run ./scripts/
  • whenever possible, translate the new strings you put in code or glade files.
  • when you change strings, please make sure that old translations get re-used.
  • commit your changes.
  • push to github.
  • open a pull request.

publishing to production

  • open the pull request page using as base a production line ghini-x.y, compared to ghini-x.y-dev.
  • make sure a bump commit is included in the differences.
  • it should be possible to automatically merge the branches.
  • create the new pull request, call it as “publish to the production line”.
  • you possibly need wait for travis-ci to perform the checks.
  • merge the changes.
  • tell the world about it: on facebook, the google group, linkedin, ...

closing step

  • review this workflow. consider this as a guideline, to yourself and to your colleagues. please help make it better and matching the practice.