This page gives an overview of the Koji code and then describes what needs to change if you want to add a new type of task. A new task could be for a new content type or assembling the results of multiple builds together, or something else that helps your workflow. New contributors to Koji should leave this page knowing where to begin and have enough understanding of Koji's architecture to be able to estimate how much work is still ahead of them.
Task Flow
A task starts with a user submitting it with the Koji client, which is a command line interface. This contacts the hub, an Apache-based server application. It leaves a row in the database that represents a "free" task, one that has not been assigned to a builder. Periodically, the builders asynchronously ping the hub asking if there are any tasks available, and at some point, one will be given the new task. The hub marks this in the database, and the builder begins executing the task (a build).
Upon completion, the builder uploads the results to the hub, including logs, binaries, environment information, and whatever else the task handler for the build dictated. The hub moves the results to a permanent shared storage solution and marks the task as completed (or failed). During this whole time, the webUI can be used to check up on progress. So the flow of work is:
Client -> Hub -> Builder -> Hub
If you wanted to add a new build type or task that was tightly integrated into Koji's data model, you would need to modify the CLI, Hub, Builder, and WebUI at a minimum. Alternatively, you could do this with a plugin, which is far simpler but less flexible.
Component Overview
Koji is comprised of several components, this section goes into details for each one, and what you potentially may need to change. Every component is written in Python, so you will need to know that language beyond a beginner level.
Koji-client
koji-client is a command line interface that provides many hooks into Koji. It allows the user to query much of the data as well as perform actions such as adding users and initiating build requests.
Option Handling
The code is in cli/koji. It uses OptionParsers extensively with interspersed arguments disabled. That means these two commands are not interpreted the same:
$ koji -u admin -p password tag-build some-tag --force some-build $ koji tag-build -u admin -p password some-tag --force some-build
The second one will generate an error, because -u and -p are not options for tag-build, they must show up before that because they are global options that can be used with any subcommand. There will be two OptionParsers used with each command. The first is used to pick up arguments to koji itself, and the second for the subcommand specified. When the first one executes (see get_options()) it will figure out the subcommand and come up with a function name based on it.
The convention is to prepend the word handle_ before it and change all hyphens to underscores. If a command does not require an account with Koji, the function handle will be prepended with anon_handle_ instead. The code will dynamically call the derived function handle which is where the second OptionParser is used to parse the remaining options. To have your code log into Koji (you're writing a handle_ function), use the activate_session function. All function signatures in the client code will get a session object, which is your interface to the hub.
Profiles
It is possible to run the Koji client with different configuration profiles so that you can interact with multiple Koji instances easily. The --profile option to the Koji command itself enables this. You should have a ~/.koji/config already, if not just copy from /etc/koji.conf to get a start. The profile command accepts an argument that matches a section in that config file. So if your config file had this:
[Fedora] authtype = ssl server = http://koji.fedoraproject.org/kojihub topdir = /mnt/koji weburl = http://koji.fedoraproject.org/koji #pkgurl = http://koji.fedoraproject.org/packages cert = ~/.fedora.cert ca = ~/.fedora-upload-ca.cert serverca = ~/.fedora-server-ca.cert [MyKoji] server = http://koji.mydomain.com/kojihub authtype = kerberos topdir = /mnt/koji weburl = http://koji.mydomain.com/koji topurl = http://download.mydomain.com/kojifiles
you could pass Fedora or MyKoji to --profile.
Creating Tasks
Once options are processed and understood, a task needs to be created on the hub so that a builder can come along and take it. This is accomplished with the maskTask method (defined on the Hub, so call it on the session object). The name of the task should match the name given to the task handler in the builder, which is explained later on.
Be sure to process the channel, priority, background, and watch/nowatch parameters too, which should be available to most new tasks. They'll be buried in the first argument to your handler function, which captures the options passed to the base Koji command.
If the client needs to make locally-available artifacts (config files, sources, kickstarts) accessible to the builder, it must be uploaded to the hub. This is the case with uploading SRPMs or kickstarts. You can easily upload this content with the session.uploadWrapper method. You can create progress bars as necessary with this snippet:
if _running_in_bg() or task_opts.noprogress:
callback = None
else:
callback = _progress_callback
serverdir = _unique_path('cli-image') # create a unique path on the hub
session.uploadWrapper(somefile, serverdir, callback=callback)
Task Arguments
If you define a new task for Koji, you'll want the task submission output to have the options ordered usefully. This output is automatically generated, but sometimes it does not capture the more important arguments you want to be displayed.
Created task 10001810 Watching tasks (this may be safely interrupted)... 10001810 thing (noarch): free 10001810 thing (noarch): free -> closed 0 free 0 open 1 done 0 failed 10001810 thing (noarch) completed successfully
In this (fake) example, you can see that "noarch" is the only option being displayed, but maybe you want something more than just the task architecture displayed, like some other options that were passed in. You can fix this behavior in koji/__init__.py in the _taskLabel function. Here you can define the string(s) to display when Koji receives status on a task. That is the return value.
Koji-Hub
koji-hub is the center of all Koji operations. It is an XML-RPC server running under mod_wsgi in Apache. koji-hub is passive in that it only receives XML-RPC calls and relies upon the build daemons and other components to initiate communication. koji-hub is the only component that has direct access to the database and is one of the two components that have to write access to the file system. If you want to make changes to the webUI (new pages or themes), you are looking in the wrong section, there is a separate component for that.
Implementation Details
The hub/kojihub.py file is where the server-side code lives. If you need to fix any server problems or want to add any new tasks, you will need to modify this file. Changes to the database schema will almost certainly require code changes too. This file gets deployed to /usr/share/koji-hub/kojihub.py, whenever you make changes to that remember to restart httpd. Also, there are cases where httpd looks for an existing .pyc file and takes it as-is, instead of re-compiling it when the code is changed.
In the code there are two large classes: RootExports and HostExports. RootExports exposes methods using XMLRPC for any client that connects to the server. The Koji CLI makes use of this quite a bit. If you want to expose a new API to any remote system, add your code here. The HostExports class does the same thing except it will ensure the requests are only coming from builders. Attempting to use an API exposed here with the CLI will fail. If your work requires the builders to call a new API, you should implement it here. Any other function defined in this file is inaccessible by remote hosts. It is generally a good practice to have the exposed APIs do very little work, and pass off control to internal functions to do the heavy lifting.
Database Interactions
Database interactions are done with raw query strings, not with any kind of modern ORM. Consider using context objects from the Koji contexts library for thread-safe interactions. The database schema is captured in the docs directory in the root of a git clone. A visualization of the schema is not available at the time of this writing.
If you plan to introduce schema changes, please update both schema.sql and provide a migration script if necessary.
Troubleshooting
The hub runs in an Apache service, so you will need to look in Apache logs for error messages if you are encountering 500 errors or the service is failing to start. Specifically, you want to check in:
- /var/log/httpd/error_log
- /var/log/httpd/ssl_error_log
If you need more specific tracebacks and debugging data, consider changing the debugging setting in /etc/koji-hub/hub.conf. Be advised the hub is very verbose with this setting on, your logs will take up gigabytes of space within several days.
Kojid
kojid is the build daemon that runs on each of the build machines. Its primary responsibility is polling for incoming build requests and handling them accordingly. Essentially kojid asks koji-hub for work. Koji also has support for tasks other than building. Creating install images is one example. kojid is responsible for handling these tasks as well. kojid uses mock for building. It also creates a fresh buildroot for every build. kojid is written in Python and communicates with koji-hub via XML-RPC.
Implementation Details
The daemon runs as a service on a host that is traditionally not the same as the hub or webUI. This is a good security practice because the service runs as root, and executes untrusted code to produce builds on a regular basis. Keeping the Hub separate limits the damage a malicious package can do to the build system as a whole. For the same reason, the filesystem that the hub keeps built software on should be mounted Read-Only on the build host. It should call APIs on the hub that are exposed through the HostExports class in the hub code. Whenever the builder accepts a task, it forks a process to carry out the build.
An initscript/unit-file is available for kojid, so it can be stopped and started like a normal service. Remember to do this when you deploy changes!
TaskHandlers
All tasks in kojid have a TaskHandler class that defines what to do when the task is picked up from the hub. The base class is defined in koji/tasks.py where a lot of useful utility methods are available. An example is uploadFile, which is used to upload logs and built binaries from a completed build to the hub since the shared filesystem is read-only.
The daemon code lives in builder/kojid, which is deployed to /usr/sbin/kojid. In there you'll notice that each task handler class has a Methods member and _taskWeight member. These must be defined, and the former is used to match the name of a waiting task (on the hub) with the task handler code to execute. Each task handler object must have a handler method defined, which is the entry point for the forked process when a builder accepts a task.
Tasks can have subtasks, which is a typical model when a build can be run on multiple architectures. In this case, developers should write 2 task handlers: one handles the build for exact one architecture, and one that assembles the results of those tasks into a single build, and sends status information to the hub. You can think of the latter handler as the parent task.
All task handler objects have a session object defined, which is the interface to use for communications with the hub. So, parent tasks should kick off child tasks using the session object's subtask method (which is part of HostExports). It should then call self.wait with all=True to wait for the results of the child tasks.
Here's a stub of what a new build task might look like:
class BuildThingTask(BaseTaskHandler):
Methods = ['thing']
_taskWeight = 0.5
def handler(self, a, b, arches, options):
subtasks = {}
for arch in arches:
subtasks[arch] = session.host.subtask(method='thingArch', a, b, arch)
results = self.wait(subtasks.values(), all=True)
# parse results and put rows in database
# put files in their final resting place
return 'Build successful'
class BuildThingArchTask(BaseTaskHandler):
Methods = ['thingArch']
_taskWeight = 2.0
def handler(self, a, b, arch):
# do the build, capture results in a variable
self.uploadFile('/path/to/some/log')
self.uploadFile('/path/to/binary/file')
return result
Source Control Managers
If you your build needs to check out code from a Source Control Manager (SCM) such as git or subversion, you can use SCM objects defined in koji/daemon.py. They take a specially formed URL as an argument to the constructor. Here's an example use. The second line is important, it makes sure the SCM is in the whitelist of SCMs allowed in /etc/kojid/kojid.conf.
scm = SCM(url)
scm.assert_allowed(self.options.allowed_scms)
directory = scm.checkout('/checkout/path', session, uploaddir, logfile)
Checking out takes 4 arguments: where to checkout, a session object (which is how authentication is handled), a directory to upload the log to, and a string representing the log file name. Using this method Koji will checkout (or clone) a remote repository and upload a log of the standard output to the task results.
Build Root Objects
It is encouraged to build software in mock chroots if appropriate. That way Koji can easily track precise details about the environment in which the build was executed. In builder/kojid a BuildRoot class is defined, which provides an interface to execute mock commands. Here's an example of their use:
broot = BuildRoot(self.session, self.options, build_tag, arch, self.id)
A session object, task options, and a build tag should be passed in as-is. You should also specify the architecture and the task ID. If you ever need to pass in specialized options to mock, look in the ImageTask.makeImgBuildRoot method to see how they are defined and passed in to the BuildRoot constructor.
Troubleshooting
The daemon writes a log file to /var/log/kojid.log. Debugging output can be turned on in /etc/kojid/kojid.conf.
Koji-Web
koji-web is a set of scripts that run in mod_wsgi and use the Cheetah templating engine to provide a web interface to Koji. It acts as a client to koji-hub providing a visual interface to perform a limited amount of administration. koji-web exposes a lot of information and also provides a means for certain operations, such as canceling builds.
The web pages are derived from Cheetah templates, the syntax of which you can read up on here. These templates are the chtml files sitting in www/kojiweb. You'll notice quickly that these templates are referencing variables, but where do they come from?
The www/kojiweb/index.py file provides them. There are several functions named after the templates they support, and in each one a dictionary called values is populated. This is how data is gathered about the task, build, archive, or whatever the page is about. Take your time with taskinfo.chtml in particular, as the conditionals there have gotten quite long. If you are adding a new task to Koji, you will need to extend this at a minimum. A new type of build task would require this, and possibly another that is specific to viewing the archived information about the build. (taskinfo vs. buildinfo)
If your web page needs to display the contents of a list or dictionary, use the $printMap function to help with that. It is often sensible to define a function that easily prints options and values in a dictionary. An example of this is in taskinfo.chtml.
#def printOpts($opts) #if $opts <strong>Options:</strong><br/> $printMap($opts, ' ') #end if #end def
Finally, if you need to expand the drop-down menus of "method" types when searching for tasks in the WebUI, you will need to add them to the _TASKS list in www/kojiweb/index.py. Add values where appropriate to _TOPLEVEL_TASKS and _PARENT_TASKS as well so that parent-child relationships show up correctly too.
Remember whenever you update a template or index.py, you will need to deploy and restart apache/httpd!
Troubleshooting
Like the hub, this component is backed by Apache, so you should follow the same techniques for debugging Koji-Web as Koji-Hub.
Kojira
kojira is a daemon that keeps the build root repodata updated. It is responsible for removing redundant build roots and cleaning up after a build request is completed.
Building and Deploying Changes
The root of the git clone for Koji code contains a Makefile that has a few targets to make building and deployment a little easier. Among them are:
- tarball: create a bz2 tarball that could be consumed in an rpm build
- rpm: create Koji rpms. The NVRs will be defined by the spec file, which is also in the same directory. The results will appear in a
noarchdirectory. - test-rpm: like rpm, but append the Release field with a date and time stamp for easy upgrade-deployment
Plugins
This section is copied from the docs/Writing_a_plugin.md file.
Koji supports different types of plugins, three of which are captured here. Depending on what you are trying to do, there are different ways to write a Koji plugin.
Koji Builder Plugins
Koji can do several things, for example, build RPMs, or live CDs. Those are types of tasks which Koji knows about. If you need to do something which Koji does not know yet how to do, you could create a Koji Builder plugin. Such a plugin would minimally look like this:
from koji.tasks import BaseTaskHandler
class MyTask(BaseTaskHandler):
Methods = ['mytask']
_taskWeight = 2.0
def handler(self, arg1, arg2, kwarg1=None):
self.logger.debug("Running my task...")
# Here is where you actually do something
A few explanations on what goes on here:
- Your task needs to inherit from
koji.tasks.BaseTaskHandler - Your task must have a
Methodsattribute, which is a list of the method names your task can handle. - You can specify the weight of your task with the
_taskWeightattribute. The more intensive (CPU, IO, ...) your task is, the higher this number should be. - The task object has a
loggerattribute, which is a Python logger with the usualdebug,info,warninganderrormethods. The messages you send with it will end up in the Koji Builder log. - Your task must have a
handler()method. That is the method Koji will call to run your task. It is the method that should actually do what you need. It can have as many positional and named arguments as you want.
Save your plugin as e.g mytask.py, then install it in the Koji Builder
plugins folder: /usr/lib/koji-builder-plugins/. Finally, edit the Koji Builder config file, /etc/kojid/kojid.conf:
# A space-separated list of plugins to enable plugins = mytask
Restart the Koji Builder service, and your plugin will be enabled. You can try running a task from your new task type with the command-line: $ koji make-task mytask arg1 arg2 kwarg1
Hub Plugins
Koji clients talk to the Koji Hub via an XMLRPC API. It is sometimes desirable to add to that API, so that clients can request things Koji does not expose right now. Such a plugin would minimally look like this:
def mymethod(arg1, arg2, kwarg1=None):
# Here is where you actually do something
mymethod.exported = True
What's happening?
- Your plugin is just a method, with whatever positional and/or named arguments you need.
- You must export your method by setting its
exportedattribute toTrue - The
context.session.assertPerm()is how you ensure that the correct permissions are available.
Save your plugin as e.g mymethod.py, then install it in the Koji Hub plugins folder, which is /usr/lib/koji-hub-plugins/
Finally, edit the Koji Hub config file, /etc/koji-hub/hub.conf:
# A space-separated list of plugins to enable Plugins = mymethod
Restart the Koji Hub service, and your plugin will be enabled. You can try calling the new XMLRPC API with the Python client library:
>>> import koji
>>> session = koji.ClientSession("http://koji/example.org/kojihub")
>>> session.mymethod(arg1, arg2, kwarg1='some value')
If you want your new XMLRPC API to require specific permissions from the user, all you need to do is add the following to your method:
from koji.context import context
def mymethod(arg1, arg2, kwarg1=None):
context.session.assertPerm("admin")
# Here is where you actually do something
mymethod.exported = True
In the example above, Koji will ensure that the user is an administrator. You could, of course, create your own permission, and check for that.
Event Plugin
You might want to run something automatically when something else happens in Koji. A typical example is to automatically sign a package right after a build finished. Another would be to send a notification to a message bus after any kind of event.
This can be achieved with a plugin too, which would look minimally as follows:
from koji.plugin import callback
@callback('preTag', 'postTag')
def mycallback(cbtype, tag, build, user, force=False):
# Here is where you actually do something
So what is this doing?
- The
@callbackdecorator allows you to declare which events should trigger your function. You can pass as many as you want. For a list of supported events, seekoji/plugins.py. - The arguments of the function depend on the event you subscribed to. As a result, you need to know how it will be called by Koji. You probably should use
*kwargsto be safe. You can see how callbacks are called in thehub/kojihub.pyfile, search for calls of therun_callbacksfunction.
Save your plugin as e.g mycallback.py, then install it in the Koji Hub plugins folder: /usr/lib/koji-hub-plugins
Finally, edit the Koji Hub config file, /etc/koji-hub/hub.conf:
# A space-separated list of plugins to enable Plugins = mycallback
Restart the Koji Hub service, and your plugin will be enabled. You can try triggering your callback plugin with the command-line. For example, if you registered a callback for the postTag event, try tagging a build: $ koji tag-build mytag mypkg-1.0-1
Patch Review
If you have a patch to submit, please send it to koji-devel@lists.fedoraproject.org. Here are some guidelines on producing preferable patches.
- Please do not "patch bomb". Keep them manageable, perhaps providing a patch for each component that is getting changes.
- Adhere to [PEP8]
- Provide patches a feature at a time, not a pile of several. Please follow the same practice for bugs.
- Consider using
git-send-email, or just attach the patches to the email - Patches should apply cleanly to HEAD of the intended branch
- Please maintain backward-compatibility up to RHEL 5 (which means Python 2.4)
There are only 3-5 people with commit access that regularly watch the mailing list, so responses may take a couple days if we're all occupied.