Skip to content
This repository has been archived by the owner. It is now read-only.
Python re-implementation of the graphwalker testing tool
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.
debian Update debian packaging Mar 13, 2015
LICENSE.txt Initial release Mar 19, 2013
setup.cfg Add flake8 config and fix some style issues Nov 24, 2015

Python Graphwalker


Python-Graphwalker is a tool for testing based on finite state machine graphs. Graphwalker reads FSMs specified by graphs, plans paths, calls model methods by name from graph labels and reports progress and results.

While conceptually derived from the Graphwalker projectgw, (implemented in java) this is a complete reimplementation from that initial concept.

Notably, there are a few differences:

  • In the original, nodes are considered states to be verified and edges actions to be taken, but this version has no ambition to enforce this convention in any way, even though it is quite useful.

  • Python Graphwalker does not understand extended FSM labels. It should ignore them, but proceed at your own risk until this is definitively dealt with one way or the other.

  • Python Graphwalker is quite promiscuous about letting you load and combine code to implement the different components of the design. Some combinations don't make sense.

  • Instead of the SWITCH_MODEL keywords in the original, this version simply allows you to load multiple graphs and stitches them together where id:s and labels match.

Generally, these simplifications makes the resulting graph much easier to reason about.

The graph for the self-test of the Interactive planner.

Overall design

The idea that has driven the design is that the graph-problems are quite orthogonal to the testing actions and that the problem of reporting the results are orthogonal to both. The graph-problems are further decomposable into path planning, stop conditions and of course loading graph files.

The added feature request to be able save and replay the path of a run dissolve into the path-recorder reporting class and the plain text graph loader.

The design is separated into these parts:

  • Model, (normally) supplied by the user as a graph file.

  • Stop condition, which bool-converts to true if its conditions are met.

  • Planner, which uses the model and stop condition to provide an iterable of plan steps as (id, name, ...) tuples.

  • Reporter, which is called on execution events.

  • Taps, installed by the reporter system to capture side-effects. (currently stdout/stderr and logging)

  • Actor supplied by the user as an object with function attributes, normally an object instance.

  • Executor that, for each step in the plan, calls the reporter and looks up and calls the named method on the actor. In addition to the step methods, it also calls a few other methods, if present on the actor.

Code loader

There is a common code-loader interface, so it's easy to load custom code and supply arguments (if any, if callable) from the command line:

  • --foo=module.module

  • --foo=module.module.function

  • --foo=module.module.class:argument,...,keyword=value,...

If the object found is callable, it will be called, with any arguments supplied, and the result used.

Model combining

Models can now be broken down into sub-graphs that can be combined at loading time. In order for this to make sense, some nodes (and optionally edges) need to have the same id and label, and these will be considered the same in the new graph. They must match on both id and label, but they are permitted to have different attributes like weight, as long as those attributes don't conflict.

Graph formats

Currently, Python Graphwalker understands a few simple file formats:


Graphs for the original Graphwalker are typically drawn using [yEd], which normally produces graphml files, so support for these have been a priority.


Plain graphviz files can also be written, which turns out to be useful: The Cartographer reporter uses dot to generate highlighted maps as it goes.

plain text

Plain text word lists are interpreted as a linear list of nodes to visit. Comments of the familiar "/* ... */" form are respected, as are line comments of both the "#" and "//" varieties. If the first node isn't labeled "Start", such a node is added.

other formats

Other formats are easy enough to add. All that you need to supply for a reader is an iterable of vertex (id, label) pairs and an iterable of (id, label, from-id, to-id) quadruples. Graphwalker will convert these to its internal formats. For write-support, you need to take a similar pair of sequences, but with the difference that for the vertex and edge tuples might be longer.


The steps to be executed by the executor are determined by one or more planners. Normally, planners are expected to examine the supplied graph and plan a traversal of it, but the lack of enforcement creates a few special opportunities.

Planners are instantiated through the common code-loader interface, so it's easy to plug in your own planner. They're called with a graph and a StopCond instance to supply an iterable containing tuples of at least two elements, as the executor expects id and label.

To generate repeatable plans, use the seed keyword argument as planners keep their own random number generators.


The simplest planner, Random, traverses the graph by randomly choosing an edge and visiting that edge and the target vertex until the StopCond is satisfied. It does not check the StopCond between edge and vertex. Edges may be weighted to skew the edge choices, by adding attributes like "weight=0.3" to a second line of edge labels. If used, weights should sum to 1.0. If only some edges have weights, the remaining edges will share the remaining weight equally.


graphwalker --stopcond=Coverage --planner=Random:seed=1337


To visit specific vertices, name them as arguments to the Goto planner. In addition to names and ids, 'random' will pick a vertex at random. If there is more than one candidate, the one closest to the current vertex will be chosen. (So this does not, currently, minimize the total path.)

An integer for the keyword argument 'repeat' will repeat the name list. (but not, nota bene, the specific vertices.) A repeat of zero will be taken to mean infinity.


graphwalker --planner=Goto:happy,random,sad,repeat=10


To visit all edges in the graph most efficiently, we'd like to generate an Eulerian trail. Since the graph is not necessarily even (semi-)Eulerian, the Euler planner copies the graph and modifies it. First, by cutting out the forced steps from the Start vertex source subgraph. The graph is then 'eulerized' by adding edges to make it Eulerian. (in-degree equal to out-degree for all vertices) After the plan is created it run through the StopCond, to get rid of extraneous steps at the end.


graphwalker --planner=Euler


There's often a wish to choose paths as the test is running when developing or debugging models. When run, Interactive lists the edges of the current vertex and prompts for input. You can choose a listed edge by entering it's number, or you can use one of the special commands:

command effect
g, goto Goes to the specified vertex*
f, force Send some arbitrary name(s) as plan steps
j, jump Set some new vertex* as the current one
d, debug Enter the pdb debugger
q, quit End the plan

*: asks if there's more than one by the name given

Notes about entering the debugger

If you quit from the debugger, you quit from the whole program. Catching BdbQuit exceptions doesn't seem to work, instead, use c/continue

You can set breakpoints in, for instance, other planners, that will drop you back into the debugger after you've left it.

StopConds - When to stop

Some planners have inherent stopping conditions, others don't, so there are independent conditions that can be applied to the plans. It's up to the planner to consult them, to they don't always cut the test off optimally, or at all.


The default stop condition is coverage of 100% of edges, which means that it will signal completion when it's seen all the edges in the graph. It can also require some percentage of vertices, or some percentage of each. The percentages are given as keywords arguments named 'edges' and 'verts' or 'vertices'.


graphwalker --stopcond=Coverage:edges=100,verts=50

graphwalker --stopcond=Coverage:vertices=25


Ignoring the difference between edges and vertices, SeenSteps will simply be done when it has seen all the steps it's looking for. The steps are given as an argument list.


graphwalker --stopcond=SeenSteps:a,e_once,b


Again ignoring the difference between edges and vertices, simply counts the test steps and signals when some number of steps have been taken. The number of steps is the first argument, or the keyword argument 'steps', defaulting to 100.


graphwalker --stopcond=CountSteps:52

graphwalker --stopcond=CountSteps:steps=52


The test executor simply uses getattr to look up callables by the names supplied by the planner, so you can implement the test code as a module, a class, or, using the programmatic interface, basically any object you like.

The callables on the test object are called without arguments for now.

In addition to the labels in the graph, a few administrative methods are also called, if present:

  • setup is called at the start of the test session with a dictionary containing the other instances involved in the test: the reporter, the model, and so on. Notably, if you want to save attachments from the test methods, you should use the reporter instance here.

  • step_begin is called before each step with the step definition. The step definition is an iterable where the first is the id and the second the name of the step.

  • step_end is called before after each step like step_begin, but with the addition of a failure, usually None. If the test failed, or there was some other exception, step_end is called with that exception, typically an AssertionError. The step_end method can permit the testing to continue by returning the exact string "RECOVER".

  • teardown is called the same way as setup, at the end.


To report the results of the tests, the reporters are all called for each event, notably step_begin and step_end.


Simply print to stdout (default) or stderr, controlled with the keyword argument output. If you are using the programming interface, you can send any file-like, writable object. Note that combinations of Log and Print quickly get really confusing.


graphwalker --reporter=Print:output=stderr


Emits to the standard python logger. The name of the logger defaults to the name of the reporting module, but can be set via the keyword argument 'logger'. The level can also be set with the keyword argument 'level'. Note that combinations of Log and Print quickly get really confusing.


graphwalker --reporter=Log:logger=moo,level=WARN


The PathRecorder simply saves the plan step names to a text file, so that the run can be replicated by feeding recording to the plain-text graph reader. The directory where the file is saved defaults to '.' but can be given as the keyword argument 'path'. Likewise name defaults to the test name but can be set with the keyword argument 'name'. The 'attach' keyword argument, if set (at all) makes it try to attach it.


graphwalker --reporter=PathRecorder:path=/tmp,name=steps

graphwalker --reporter=PathRecorder:attach=true,name=steps


To map the progress of the test graphically, the Cartographer reporter emits graphviz files with the current step highlighted. The keyword arguments 'dotpath' and 'imgpath' control where the graphviz input and output files go, respectively, bot defaulting to '.'. The image type defaults to PNG but can be set using the keyword argument 'imgtype'. The 'attach' keyword argument, if set (at all) makes it try to attach it.


graphwalker --reporter=Cartographer

graphwalker --reporter=Cartographer:imgtype=jpg,attach=1

graphwalker --reporter=Cartographer:dotpath=/tmp,imgpath=./www


Currently, the there are only taps for streams and the logging system. Both the logging tap and taps of standard out & error are included by default.


Graphwalker itself needs a lot more, and a lot more devious tests.


The first iteration of the Python port of Graphwalker was written by Viktor Holmberg, Harald Hartwig and Chongyang Sun under the direction of Nils Österling (tester) and Anders Eurenius (developer).

This iteration was rewritten from scratch by Anders Eurenius to incorporate everything we learned from the first.


The license we have chosen is the Apache License, version 2.0. You should find the full text in the file named "LICENSE.txt".

You can’t perform that action at this time.