For advanced users, Buildbot acts as a framework supporting a customized build application. For the most part, such configurations consist of subclasses set up for use in a regular Buildbot configuration file.

This chapter describes some of the more common idioms in advanced Buildbot configurations.

At the moment, this chapter is an unordered set of suggestions; if you'd like to clean it up, fork the project on github and get started!

Programmatic Configuration Generation

Bearing in mind that master.cfg is a Python file, large configurations can be shortened considerably by judicious use of Python loops. For example, the following will generate a builder for each of a range of supported versions of Python:

pythons = [ 'python2.4', 'python2.5', 'python2.6', 'python2.7',
            'python3.2', python3.3' ]
pytest_slaves = [ "slave%s" % n for n in range(10) ]
for python in pythons:
    f = BuildFactory()
    f.addStep(ShellCommand(command=[ python, '' ]))
            name="test-%s" % python,

Merge Request Functions

The logic Buildbot uses to decide which build request can be merged can be customized by providing a Python function (a callable) instead of True or False described in Merging Build Requests.

The callable will be invoked with three positional arguments: a Builder object and two BuildRequest objects. It should return true if the requests can be merged, and False otherwise. For example:

def mergeRequests(builder, req1, req2):
    "any requests with the same branch can be merged"
    return req1.branch == req2.branch
c['mergeRequests'] = mergeRequests

In many cases, the details of the SourceStamps and BuildRequests are important. In this example, only BuildRequests with the same "reason" are merged; thus developers forcing builds for different reasons will see distinct builds. Note the use of the canBeMergedWith method to access the source stamp compatibility algorithm.

def mergeRequests(builder, req1, req2):
    if req1.source.canBeMergedWith(req2.source) and  req1.reason == req2.reason:
       return True
    return False
c['mergeRequests'] = mergeRequests

If it's necessary to perform some extended operation to determine whether two requests can be merged, then the mergeRequests callable may return its result via Deferred. Note, however, that the number of invocations of the callable is proportional to the square of the request queue length, so a long-running callable may cause undesirable delays when the queue length grows. For example:

def mergeRequests(builder, req1, req2):
    d = defer.gatherResults([
    def process(info1, info2):
        return info1 == info2
    return d
c['mergeRequests'] = mergeRequests

Builder Priority Functions

The prioritizeBuilders configuration key specifies a function which is called with two arguments: a BuildMaster and a list of Builder objects. It should return a list of the same Builder objects, in the desired order. It may also remove items from the list if builds should not be started on those builders. If necessary, this function can return its results via a Deferred (it is called with maybeDeferred).

A simple prioritizeBuilders implementation might look like this:

def prioritizeBuilders(buildmaster, builders):
    """Prioritize builders.  'finalRelease' builds have the highest
    priority, so they should be built before running tests, or
    creating builds."""
    builderPriorities = {
        "finalRelease": 0,
        "test": 1,
        "build": 2,
    builders.sort(key=lambda b: builderPriorities.get(, 0))
    return builders

c['prioritizeBuilders'] = prioritizeBuilders

Build Priority Functions

When a builder has multiple pending build requests, it uses a nextBuild function to decide which build it should start first. This function is given two parameters: the Builder, and a list of BuildRequest objects representing pending build requests.

A simple function to prioritize release builds over other builds might look like this:

def nextBuild(bldr, requests):
    for r in requests:
        if r.source.branch == 'release':
            return r
    return requests[0]

If some non-immediate result must be calculated, the nextBuild function can also return a Deferred:

def nextBuild(bldr, requests):
    d = get_request_priorities(requests)
    def pick(priorities):
        if requests:
            return sorted(zip(priorities, requests))[0][1]
    return d

Customizing SVNPoller

Each source file that is tracked by a Subversion repository has a fully-qualified SVN URL in the following form: ({REPOURL})({PROJECT-plus-BRANCH})({FILEPATH}). When you create the SVNPoller, you give it a svnurl value that includes all of the {REPOURL} and possibly some portion of the {PROJECT-plus-BRANCH} string. The SVNPoller is responsible for producing Changes that contain a branch name and a {FILEPATH} (which is relative to the top of a checked-out tree). The details of how these strings are split up depend upon how your repository names its branches.


One common layout is to have all the various projects that share a repository get a single top-level directory each, with branches, tags, and trunk subdirectories:


To set up a SVNPoller that watches the Amanda trunk (and nothing else), we would use the following, using the default split_file:

from buildbot.changes.svnpoller import SVNPoller
c['change_source'] = SVNPoller(

In this case, every Change that our SVNPoller produces will have its branch attribute set to None, to indicate that the Change is on the trunk. No other sub-projects or branches will be tracked.

If we want our ChangeSource to follow multiple branches, we have to do two things. First we have to change our svnurl= argument to watch more than just amanda/trunk. We will set it to amanda so that we'll see both the trunk and all the branches. Second, we have to tell SVNPoller how to split the ({PROJECT-plus-BRANCH})({FILEPATH}) strings it gets from the repository out into ({BRANCH}) and ({FILEPATH})`.

We do the latter by providing a split_file function. This function is responsible for splitting something like branches/3_3/common-src/amanda.h into branch='branches/3_3' and filepath='common-src/amanda.h'. The function is always given a string that names a file relative to the subdirectory pointed to by the SVNPoller's svnurl= argument. It is expected to return a dictionary with at least the path key. The splitter may optionally set branch, project and repository. For backwards compatibility it may return a tuple of (branchname, path). It may also return None to indicate that the file is of no interest.


the function should return branches/3_3 rather than just 3_3 because the SVN checkout step, will append the branch name to the baseURL, which requires that we keep the branches component in there. Other VC schemes use a different approach towards branches and may not require this artifact.

If your repository uses this same {PROJECT}/{BRANCH}/{FILEPATH} naming scheme, the following function will work:

def split_file_branches(path):
    pieces = path.split('/')
    if len(pieces) > 1 and pieces[0] == 'trunk':
        return (None, '/'.join(pieces[1:]))
    elif len(pieces) > 2 and pieces[0] == 'branches':
        return ('/'.join(pieces[0:2]),
        return None

In fact, this is the definition of the provided split_file_branches function. So to have our Twisted-watching SVNPoller follow multiple branches, we would use this:

from buildbot.changes.svnpoller import SVNPoller, split_file_branches
c['change_source'] = SVNPoller("svn://",

Changes for all sorts of branches (with names like "branches/1.5.x", and None to indicate the trunk) will be delivered to the Schedulers. Each Scheduler is then free to use or ignore each branch as it sees fit.

If you have multiple projects in the same repository your split function can attach a project name to the Change to help the Scheduler filter out unwanted changes:

from buildbot.changes.svnpoller import split_file_branches
def split_file_projects_branches(path):
    if not "/" in path:
        return None
    project, path = path.split("/", 1)
    f = split_file_branches(path)
    if f:
        info = dict(project=project, path=f[1])
        if f[0]:
            info['branch'] = f[0]
        return info
    return f

Again, this is provided by default. To use it you would do this:

from buildbot.changes.svnpoller import SVNPoller, split_file_projects_branches
c['change_source'] = SVNPoller(

Note here that we are monitoring at the root of the repository, and that within that repository is a amanda subdirectory which in turn has trunk and branches. It is that amanda subdirectory whose name becomes the project field of the Change.


Another common way to organize a Subversion repository is to put the branch name at the top, and the projects underneath. This is especially frequent when there are a number of related sub-projects that all get released in a group.

For example, hosts a project named Nevow as well as one named Quotient. In a checked-out Nevow tree there is a directory named formless that contains a python source file named This repository is accessible via webdav (and thus uses an http: scheme) through the hostname. There are many branches in this repository, and they use a ({BRANCHNAME})/({PROJECT}) naming policy.

The fully-qualified SVN URL for the trunk version of is The 1.5.x branch version of this file would have a URL of The whole Nevow trunk would be checked out with, while the Quotient trunk would be checked out using

Now suppose we want to have an SVNPoller that only cares about the Nevow trunk. This case looks just like the {PROJECT}/{BRANCH} layout described earlier:

from buildbot.changes.svnpoller import SVNPoller
c['change_source'] = SVNPoller("")

But what happens when we want to track multiple Nevow branches? We have to point our svnurl= high enough to see all those branches, but we also don't want to include Quotient changes (since we're only building Nevow). To accomplish this, we must rely upon the split_file function to help us tell the difference between files that belong to Nevow and those that belong to Quotient, as well as figuring out which branch each one is on.

from buildbot.changes.svnpoller import SVNPoller
c['change_source'] = SVNPoller("",

The my_file_splitter function will be called with repository-relative pathnames like:

This is a Nevow file, on the trunk. We want the Change that includes this to see a filename of formless/, and a branch of None
This is a Nevow file, on a branch. We want to get branch='branches/1.5.x' and filename='formless/'.
This is a Quotient file, so we want to ignore it by having my_file_splitter return None.
This is also a Quotient file, which should be ignored.

The following definition for my_file_splitter will do the job:

def my_file_splitter(path):
    pieces = path.split('/')
    if pieces[0] == 'trunk':
        branch = None
        pieces.pop(0) # remove 'trunk'
    elif pieces[0] == 'branches':
        pieces.pop(0) # remove 'branches'
        # grab branch name
        branch = 'branches/' + pieces.pop(0)
        return None # something weird
    projectname = pieces.pop(0)
    if projectname != 'Nevow':
        return None # wrong project
    return dict(branch=branch, path='/'.join(pieces))

If you later decide you want to get changes for Quotient as well you could replace the last 3 lines with simply:

return dict(project=projectname, branch=branch, path='/'.join(pieces))

Writing Change Sources

For some version-control systems, making Buildbot aware of new changes can be a challenge. If the pre-supplied classes in Change Sources are not sufficient, then you will need to write your own.

There are three approaches, one of which is not even a change source. The first option is to write a change source that exposes some service to which the version control system can "push" changes. This can be more complicated, since it requires implementing a new service, but delivers changes to Buildbot immediately on commit.

The second option is often preferable to the first: implement a notification service in an external process (perhaps one that is started directly by the version control system, or by an email server) and delivers changes to Buildbot via PBChangeSource. This section does not describe this particular approach, since it requires no customization within the buildmaster process.

The third option is to write a change source which polls for changes - repeatedly connecting to an external service to check for new changes. This works well in many cases, but can produce a high load on the version control system if polling is too frequent, and can take too long to notice changes if the polling is not frequent enough.

Writing a Notification-based Change Source

class buildbot.changes.base.ChangeSource

A custom change source must implement buildbot.interfaces.IChangeSource.

The easiest way to do this is to subclass buildbot.changes.base.ChangeSource, implementing the describe method to describe the instance. ChangeSource is a Twisted service, so you will need to implement the startService and stopService methods to control the means by which your change source receives notifications.

When the class does receive a change, it should call self.master.addChange(..) to submit it to the buildmaster. This method shares the same parameters as master.db.changes.addChange, so consult the API documentation for that function for details on the available arguments.

You will probably also want to set compare_attrs to the list of object attributes which Buildbot will use to compare one change source to another when reconfiguring. During reconfiguration, if the new change source is different from the old, then the old will be stopped and the new started.

Writing a Change Poller

class buildbot.changes.base.PollingChangeSource

Polling is a very common means of seeking changes, so Buildbot supplies a utility parent class to make it easier. A poller should subclass buildbot.changes.base.PollingChangeSource, which is a subclass of ChangeSource. This subclass implements the Service methods, and causes the poll method to be called every self.pollInterval seconds. This method should return a Deferred to signal its completion.

Aside from the service methods, the other concerns in the previous section apply here, too.

Writing a New Latent Buildslave Implementation

Writing a new latent buildslave should only require subclassing buildbot.buildslave.AbstractLatentBuildSlave and implementing start_instance and stop_instance.

def start_instance(self):
    # responsible for starting instance that will try to connect with this
    # master. Should return deferred. Problems should use an errback. The
    # callback value can be None, or can be an iterable of short strings to
    # include in the "substantiate success" status message, such as
    # identifying the instance that started.
    raise NotImplementedError

def stop_instance(self, fast=False):
    # responsible for shutting down instance. Return a deferred. If `fast`,
    # we're trying to shut the master down, so callback as soon as is safe.
    # Callback value is ignored.
    raise NotImplementedError

See buildbot.ec2buildslave.EC2LatentBuildSlave for an example, or see the test example buildbot.test_slaves.FakeLatentBuildSlave.

Custom Build Classes

The standard BuildFactory object creates Build objects by default. These Builds will each execute a collection of BuildSteps in a fixed sequence. Each step can affect the results of the build, but in general there is little intelligence to tie the different steps together.

By setting the factory's buildClass attribute to a different class, you can instantiate a different build class. This might be useful, for example, to create a build class that dynamically determines which steps to run. The skeleton of such a project would look like:

class DynamicBuild(Build):
    # .. override some methods

f = factory.BuildFactory()
f.buildClass = DynamicBuild

Factory Workdir Functions

It is sometimes helpful to have a build's workdir determined at runtime based on the parameters of the build. To accomplish this, set the workdir attribute of the build factory to a callable. That callable will be invoked with the SourceStamp for the build, and should return the appropriate workdir. Note that the value must be returned immediately - Deferreds are not supported.

This can be useful, for example, in scenarios with multiple repositories submitting changes to BuildBot. In this case you likely will want to have a dedicated workdir per repository, since otherwise a sourcing step with mode = "update" will fail as a workdir with a working copy of repository A can't be "updated" for changes from a repository B. Here is an example how you can achieve workdir-per-repo:

def workdir(source_stamp):
    return hashlib.md5 (source_stamp.repository).hexdigest()[:8]

build_factory = factory.BuildFactory()
build_factory.workdir = workdir

# ...
builders.append ({'name': 'mybuilder',
                  'slavename': 'myslave',
                  'builddir': 'mybuilder',
                  'factory': build_factory})

The end result is a set of workdirs like

Repo1 => <buildslave-base>/mybuilder/a78890ba
Repo2 => <buildslave-base>/mybuilder/0823ba88

You could make the workdir function compute other paths, based on parts of the repo URL in the sourcestamp, or lookup in a lookup table based on repo URL. As long as there is a permanent 1:1 mapping between repos and workdir, this will work.

Writing New BuildSteps

While it is a good idea to keep your build process self-contained in the source code tree, sometimes it is convenient to put more intelligence into your Buildbot configuration. One way to do this is to write a custom BuildStep. Once written, this Step can be used in the master.cfg file.

The best reason for writing a custom BuildStep is to better parse the results of the command being run. For example, a BuildStep that knows about JUnit could look at the logfiles to determine which tests had been run, how many passed and how many failed, and then report more detailed information than a simple rc==0 -based good/bad decision.

Buildbot has acquired a large fleet of build steps, and sports a number of knobs and hooks to make steps easier to write. This section may seem a bit overwhelming, but most custom steps will only need to apply one or two of the techniques outlined here.

For complete documentation of the build step interfaces, see BuildSteps.

Writing BuildStep Constructors

Build steps act as their own factories, so their constructors are a bit more complex than necessary. In the configuration file, a BuildStep object is instantiated, but because steps store state locally while executing, this object cannot be used during builds.

Consider the use of a BuildStep in master.cfg:

f.addStep(MyStep(someopt="stuff", anotheropt=1))

This creates a single instance of class MyStep. However, Buildbot needs a new object each time the step is executed. An instance of BuildStep remembers how it was constructed, and can create copies of itself. When writing a new step class, then, keep in mind are that you cannot do anything "interesting" in the constructor -- limit yourself to checking and storing arguments.

It is customary to call the parent class's constructor with all otherwise-unspecified keyword arguments. Keep a **kwargs argument on the end of your options, and pass that up to the parent class's constructor.

The whole thing looks like this:

class Frobnify(LoggingBuildStep):
    def __init__(self,

        # check
        if frob_how_many is None:
            raise TypeError("Frobnify argument how_many is required")

        # override a parent option
        kwargs['parentOpt'] = 'xyz'

        # call parent
        LoggingBuildStep.__init__(self, **kwargs)

        # set Frobnify attributes
        self.frob_what = frob_what
        self.frob_how_many = how_many
        self.frob_how = frob_how

class FastFrobnify(Frobnify):
    def __init__(self,
        Frobnify.__init__(self, **kwargs)
        self.speed = speed

Running Commands

To spawn a command in the buildslave, create a RemoteCommand instance in your step's start method and run it with runCommand:

cmd = RemoteCommand(args)
d = self.runCommand(cmd)

To add a LogFile, use addLog. Make sure the log gets closed when it finishes. When giving a Logfile to a RemoteShellCommand, just ask it to close the log when the command completes:

log = self.addLog('output')
cmd.useLog(log, closeWhenFinished=True)

Updating Status


Capturing Logfiles

Each BuildStep has a collection of logfiles. Each one has a short name, like stdio or warnings. Each LogFile contains an arbitrary amount of text, usually the contents of some output file generated during a build or test step, or a record of everything that was printed to stdout/stderr during the execution of some command.

These LogFiles are stored to disk, so they can be retrieved later.

Each can contain multiple channels, generally limited to three basic ones: stdout, stderr, and headers. For example, when a ShellCommand runs, it writes a few lines to the headers channel to indicate the exact argv strings being run, which directory the command is being executed in, and the contents of the current environment variables. Then, as the command runs, it adds a lot of stdout and stderr messages. When the command finishes, a final header line is added with the exit code of the process.

Status display plugins can format these different channels in different ways. For example, the web page shows LogFiles as text/html, with header lines in blue text, stdout in black, and stderr in red. A different URL is available which provides a text/plain format, in which stdout and stderr are collapsed together, and header lines are stripped completely. This latter option makes it easy to save the results to a file and run grep or whatever against the output.

Each BuildStep contains a mapping (implemented in a python dictionary) from LogFile name to the actual LogFile objects. Status plugins can get a list of LogFiles to display, for example, a list of HREF links that, when clicked, provide the full contents of the LogFile.

Using LogFiles in custom BuildSteps

The most common way for a custom BuildStep to use a LogFile is to summarize the results of a ShellCommand (after the command has finished running). For example, a compile step with thousands of lines of output might want to create a summary of just the warning messages. If you were doing this from a shell, you would use something like:

grep "warning:" output.log >warnings.log

In a custom BuildStep, you could instead create a warnings LogFile that contained the same text. To do this, you would add code to your createSummary method that pulls lines from the main output log and creates a new LogFile with the results:

def createSummary(self, log):
    warnings = []
    sio = StringIO.StringIO(log.getText())
    for line in sio.readlines():
        if "warning:" in line:
    self.addCompleteLog('warnings', "".join(warnings))

This example uses the addCompleteLog method, which creates a new LogFile, puts some text in it, and then closes it, meaning that no further contents will be added. This LogFile will appear in the HTML display under an HREF with the name warnings, since that is the name of the LogFile.

You can also use addHTMLLog to create a complete (closed) LogFile that contains HTML instead of plain text. The normal LogFile will be HTML-escaped if presented through a web page, but the HTML LogFile will not. At the moment this is only used to present a pretty HTML representation of an otherwise ugly exception traceback when something goes badly wrong during the BuildStep.

In contrast, you might want to create a new LogFile at the beginning of the step, and add text to it as the command runs. You can create the LogFile and attach it to the build by calling addLog, which returns the LogFile object. You then add text to this LogFile by calling methods like addStdout and addHeader. When you are done, you must call the finish method so the LogFile can be closed. It may be useful to create and populate a LogFile like this from a LogObserver method - see Adding LogObservers.

The logfiles= argument to ShellCommand (see ShellCommand) creates new LogFiles and fills them in realtime by asking the buildslave to watch a actual file on disk. The buildslave will look for additions in the target file and report them back to the BuildStep. These additions will be added to the LogFile by calling addStdout. These secondary LogFiles can be used as the source of a LogObserver just like the normal stdio LogFile.

Reading Logfiles

Once a LogFile has been added to a BuildStep with addLog, addCompleteLog, addHTMLLog, or logfiles={}, your BuildStep can retrieve it by using getLog:

class MyBuildStep(ShellCommand):
    logfiles = @{ "nodelog": "_test/node.log" @}

    def evaluateCommand(self, cmd):
        nodelog = self.getLog("nodelog")
        if "STARTED" in nodelog.getText():
            return SUCCESS
            return FAILURE

Adding LogObservers

Most shell commands emit messages to stdout or stderr as they operate, especially if you ask them nicely with a --verbose flag of some sort. They may also write text to a log file while they run. Your BuildStep can watch this output as it arrives, to keep track of how much progress the command has made. You can get a better measure of progress by counting the number of source files compiled or test cases run than by merely tracking the number of bytes that have been written to stdout. This improves the accuracy and the smoothness of the ETA display.

To accomplish this, you will need to attach a LogObserver to one of the log channels, most commonly to the stdio channel but perhaps to another one which tracks a log file. This observer is given all text as it is emitted from the command, and has the opportunity to parse that output incrementally. Once the observer has decided that some event has occurred (like a source file being compiled), it can use the setProgress method to tell the BuildStep about the progress that this event represents.

There are a number of pre-built LogObserver classes that you can choose from (defined in buildbot.process.buildstep, and of course you can subclass them to add further customization. The LogLineObserver class handles the grunt work of buffering and scanning for end-of-line delimiters, allowing your parser to operate on complete stdout/stderr lines. (Lines longer than a set maximum length are dropped; the maximum defaults to 16384 bytes, but you can change it by calling setMaxLineLength on your LogLineObserver instance. Use sys.maxint for effective infinity.)

For example, let's take a look at the TrialTestCaseCounter, which is used by the Trial step to count test cases as they are run. As Trial executes, it emits lines like the following:

buildbot.test.test_config.ConfigTest.testDebugPassword ... [OK]
buildbot.test.test_config.ConfigTest.testEmpty ... [OK]
buildbot.test.test_config.ConfigTest.testIRC ... [FAIL]
buildbot.test.test_config.ConfigTest.testLocks ... [OK]

When the tests are finished, trial emits a long line of ====== and then some lines which summarize the tests that failed. We want to avoid parsing these trailing lines, because their format is less well-defined than the [OK] lines.

The parser class looks like this:

from buildbot.process.buildstep import LogLineObserver

class TrialTestCaseCounter(LogLineObserver):
    _line_re = re.compile(r'^([\w\.]+) \.\.\. \[([^\]]+)\]$')
    numTests = 0
    finished = False

    def outLineReceived(self, line):
        if self.finished:
        if line.startswith("=" * 40):
            self.finished = True

        m =
        if m:
            testname, result = m.groups()
            self.numTests += 1
            self.step.setProgress('tests', self.numTests)

This parser only pays attention to stdout, since that's where trial writes the progress lines. It has a mode flag named finished to ignore everything after the ==== marker, and a scary-looking regular expression to match each line while hopefully ignoring other messages that might get displayed as the test runs.

Each time it identifies a test has been completed, it increments its counter and delivers the new progress value to the step with @code{self.step.setProgress}. This class is specifically measuring progress along the tests metric, in units of test cases (as opposed to other kinds of progress like the output metric, which measures in units of bytes). The Progress-tracking code uses each progress metric separately to come up with an overall completion percentage and an ETA value.

To connect this parser into the Trial build step, Trial.__init__ ends with the following clause:

# this counter will feed Progress along the 'test cases' metric
counter = TrialTestCaseCounter()
self.addLogObserver('stdio', counter)
self.progressMetrics += ('tests',)

This creates a TrialTestCaseCounter and tells the step that the counter wants to watch the stdio log. The observer is automatically given a reference to the step in its step attribute.

Using Properties

In custom BuildSteps, you can get and set the build properties with the getProperty/setProperty methods. Each takes a string for the name of the property, and returns or accepts an arbitrary object. For example:

class MakeTarball(ShellCommand):
    def start(self):
        if self.getProperty("os") == "win":
            self.setCommand([ ... ]) # windows-only command
            self.setCommand([ ... ]) # equivalent for other systems

Remember that properties set in a step may not be available until the next step begins. In particular, any Property or Interpolate instances for the current step are interpolated before the start method begins.

BuildStep URLs

Each BuildStep has a collection of links. Like its collection of LogFiles, each link has a name and a target URL. The web status page creates HREFs for each link in the same box as it does for LogFiles, except that the target of the link is the external URL instead of an internal link to a page that shows the contents of the LogFile.

These external links can be used to point at build information hosted on other servers. For example, the test process might produce an intricate description of which tests passed and failed, or some sort of code coverage data in HTML form, or a PNG or GIF image with a graph of memory usage over time. The external link can provide an easy way for users to navigate from the buildbot's status page to these external web sites or file servers. Note that the step itself is responsible for insuring that there will be a document available at the given URL (perhaps by using scp to copy the HTML output to a ~/public_html/ directory on a remote web server). Calling addURL does not magically populate a web server.

To set one of these links, the BuildStep should call the addURL method with the name of the link and the target URL. Multiple URLs can be set.

In this example, we assume that the make test command causes a collection of HTML files to be created and put somewhere on the web server, in a filename that incorporates the build number.

class TestWithCodeCoverage(BuildStep):
    command = ["make", "test",

    def createSummary(self, log):
        buildnumber = self.getProperty("buildnumber")
        url = "" % buildnumber
        self.addURL("coverage", url)

You might also want to extract the URL from some special message output by the build process itself:

class TestWithCodeCoverage(BuildStep):
    command = ["make", "test",

    def createSummary(self, log):
        output = StringIO(log.getText())
        for line in output.readlines():
            if line.startswith("coverage-url:"):
                url = line[len("coverage-url:"):].strip()
                self.addURL("coverage", url)

Note that a build process which emits both stdout and stderr might cause this line to be split or interleaved between other lines. It might be necessary to restrict the getText call to only stdout with something like this:

output = StringIO("".join([c[1]
                           for c in log.getChunks()
                           if c[0] == LOG_CHANNEL_STDOUT]))

Of course if the build is run under a PTY, then stdout and stderr will be merged before the buildbot ever sees them, so such interleaving will be unavoidable.

A Somewhat Whimsical Example

Let's say that we've got some snazzy new unit-test framework called Framboozle. It's the hottest thing since sliced bread. It slices, it dices, it runs unit tests like there's no tomorrow. Plus if your unit tests fail, you can use its name for a Web 2.1 startup company, make millions of dollars, and hire engineers to fix the bugs for you, while you spend your afternoons lazily hang-gliding along a scenic pacific beach, blissfully unconcerned about the state of your tests. [1]

To run a Framboozle-enabled test suite, you just run the 'framboozler' command from the top of your source code tree. The 'framboozler' command emits a bunch of stuff to stdout, but the most interesting bit is that it emits the line "FNURRRGH!" every time it finishes running a test case You'd like to have a test-case counting LogObserver that watches for these lines and counts them, because counting them will help the buildbot more accurately calculate how long the build will take, and this will let you know exactly how long you can sneak out of the office for your hang-gliding lessons without anyone noticing that you're gone.

This will involve writing a new BuildStep (probably named "Framboozle") which inherits from ShellCommand. The BuildStep class definition itself will look something like this:

from import ShellCommand
from buildbot.process.buildstep import LogLineObserver

class FNURRRGHCounter(LogLineObserver):
    numTests = 0
    def outLineReceived(self, line):
        if "FNURRRGH!" in line:
            self.numTests += 1
            self.step.setProgress('tests', self.numTests)

class Framboozle(ShellCommand):
    command = ["framboozler"]

    def __init__(self, **kwargs):
        ShellCommand.__init__(self, **kwargs)   # always upcall!
        counter = FNURRRGHCounter())
        self.addLogObserver('stdio', counter)
        self.progressMetrics += ('tests',)

So that's the code that we want to wind up using. How do we actually deploy it?

You have a couple of different options.

Option 1: The simplest technique is to simply put this text (everything from START to FINISH) in your master.cfg file, somewhere before the BuildFactory definition where you actually use it in a clause like:

f = BuildFactory()

Remember that master.cfg is secretly just a python program with one job: populating the BuildmasterConfig dictionary. And python programs are allowed to define as many classes as they like. So you can define classes and use them in the same file, just as long as the class is defined before some other code tries to use it.

This is easy, and it keeps the point of definition very close to the point of use, and whoever replaces you after that unfortunate hang-gliding accident will appreciate being able to easily figure out what the heck this stupid "Framboozle" step is doing anyways. The downside is that every time you reload the config file, the Framboozle class will get redefined, which means that the buildmaster will think that you've reconfigured all the Builders that use it, even though nothing changed. Bleh.

Option 2: Instead, we can put this code in a separate file, and import it into the master.cfg file just like we would the normal buildsteps like ShellCommand and SVN.

Create a directory named ~/lib/python, put everything from START to FINISH in ~/lib/python/, and run your buildmaster using:

PYTHONPATH=~/lib/python buildbot start MASTERDIR

or use the Makefile.buildbot to control the way buildbot start works. Or add something like this to something like your ~/.bashrc or ~/.bash_profile or ~/.cshrc:

export PYTHONPATH=~/lib/python

Once we've done this, our master.cfg can look like:

from framboozle import Framboozle
f = BuildFactory()


import framboozle
f = BuildFactory()

(check out the python docs for details about how "import" and "from A import B" work).

What we've done here is to tell python that every time it handles an "import" statement for some named module, it should look in our ~/lib/python/ for that module before it looks anywhere else. After our directories, it will try in a bunch of standard directories too (including the one where buildbot is installed). By setting the PYTHONPATH environment variable, you can add directories to the front of this search list.

Python knows that once it "import"s a file, it doesn't need to re-import it again. This means that reconfiguring the buildmaster (with buildbot reconfig, for example) won't make it think the Framboozle class has changed every time, so the Builders that use it will not be spuriously restarted. On the other hand, you either have to start your buildmaster in a slightly weird way, or you have to modify your environment to set the PYTHONPATH variable.

Option 3: Install this code into a standard python library directory

Find out what your python's standard include path is by asking it:

80:warner@luther% python
Python 2.4.4c0 (#2, Oct  2 2006, 00:57:46)
[GCC 4.1.2 20060928 (prerelease) (Debian 4.1.1-15)] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import sys
>>> import pprint
>>> pprint.pprint(sys.path)

In this case, putting the code into /usr/local/lib/python2.4/site-packages/ would work just fine. We can use the same master.cfg import framboozle statement as in Option 2. By putting it in a standard include directory (instead of the decidedly non-standard ~/lib/python), we don't even have to set PYTHONPATH to anything special. The downside is that you probably have to be root to write to one of those standard include directories.

Option 4: Submit the code for inclusion in the Buildbot distribution

Make a fork of buildbot on or post a patch in a bug at In either case, post a note about your patch to the mailing list, so others can provide feedback and, eventually, commit it.

from buildbot.steps import framboozle f = BuildFactory() f.addStep(SVN(svnurl="stuff")) f.addStep(framboozle.Framboozle())

And then you don't even have to install anywhere on your system, since it will ship with Buildbot. You don't have to be root, you don't have to set PYTHONPATH. But you do have to make a good case for Framboozle being worth going into the main distribution, you'll probably have to provide docs and some unit test cases, you'll need to figure out what kind of beer the author likes (IPA's and Stouts for Dustin), and then you'll have to wait until the next release. But in some environments, all this is easier than getting root on your buildmaster box, so the tradeoffs may actually be worth it.

Putting the code in master.cfg (1) makes it available to that buildmaster instance. Putting it in a file in a personal library directory (2) makes it available for any buildmasters you might be running. Putting it in a file in a system-wide shared library directory (3) makes it available for any buildmasters that anyone on that system might be running. Getting it into the buildbot's upstream repository (4) makes it available for any buildmasters that anyone in the world might be running. It's all a matter of how widely you want to deploy that new class.

Writing New Status Plugins

Each status plugin is an object which provides the twisted.application.service.IService interface, which creates a tree of Services with the buildmaster at the top [not strictly true]. The status plugins are all children of an object which implements buildbot.interfaces.IStatus, the main status object. From this object, the plugin can retrieve anything it wants about current and past builds. It can also subscribe to hear about new and upcoming builds.

Status plugins which only react to human queries (like the Waterfall display) never need to subscribe to anything: they are idle until someone asks a question, then wake up and extract the information they need to answer it, then they go back to sleep. Plugins which need to act spontaneously when builds complete (like the MailNotifier plugin) need to subscribe to hear about new builds.

If the status plugin needs to run network services (like the HTTP server used by the Waterfall plugin), they can be attached as Service children of the plugin itself, using the IServiceCollection interface.

[1] is still available. Remember, I get 10% :).