Vixie Cron How I Loath Thee.....

As some know, I work as the senior systems administrator for a company in the southeast SF Bay Area. Recently, we had to move to a new security system, after having been broken into four times in the last _four_ years.

One of the features of this new system is that it ties into our phone system to alert the building that the alarm is about to be armed. To do this, a dial out to a special phone number needs to occur at the desired hour and then voila! Instant Alert throughout our suite! So, we've setup an email to fax service that we can easily trigger by a simple email.

So I wrote a REALLY simple shell script shown below (with the fax and email addresses elided):

#!/bin/bash

set -e

echo "Sending Security Page NOW!"

/usr/sbin/sendmail -t <<EOF
FROM: SOMERANDOMSMUCK@minervanetworks.com
TO:   1234567890@fax.com
CC:   SOMERANDOMSMUCK@minervanetworks.com
SUBJECT: Alarm Set

Set.

EOF

RET=$?
if [ $? -eq 0 ]; then
    echo "SUCCESS"
fi

As I said, simple.

From there I setup my crontab on one of the Linux box in my cube to have it go off at the apropriote time, 7:55PM Monday through Friday:

MAILTO=SOMERANDOMSMUCK@minervanetworks.com

50 19 * * 1-5 /usr/bin/securityfax.sh

Tested out the script, thinking that "Hey, I'm all done". Wrong.

I come to find out that Vixie Cron can't seem to tell time! I get the alert emails at 8:31PM from the system!

At this point, looks like I'll be replacing Vixie Cron on my workstation (CentOS 5). Can anyone advise a good Cron Daemon replacement that has enough similarities to vixie cron that I won't have to relearn too much and can actually interpret time settings in the crontab as more than mere suggestions?

What I've Been Doing Lately... Or Better Said, Coding Lately.

Lately I've been doing research to get an Exchange resource working and usable for KDE. During this time, I've had to have a working trunk build since OpenSuSE's Build Service tends to have custom patches for their build of KDE that make some things not behave or look how the default KDE trunk tree acts/looks (that's not a comment that that's bad, just not my want for when I do development, but I digress...)

To facilitate this, I'd been using the shell scripts that KDE has on Techbase and on the OpenSuSE wiki to build trunk. I've found however, that my time would be better spent working on code, instead of waiting on code to compile when at home. Since I wanted something that would even fully track Qt's Git tree and email me when builds ran or error-ed out, kdesvn-build was not going to meet my needs. What came out of it was a hacky little Perl script called KDEBuilder that was fairly fragile and frankly not designed at all.

While thinking over the mess that was KDEBuilder, I realized that what I *really* wanted was something that was flexible, feature rich, and wasn't an overblown ported shell script on steroids. Out of this came the design for Direktor, a continuous build system written in a clean Perl code style (heavily influenced by the Google Perl Style guide....) To better describe Direktor, I've pasted the ReadMe.txt from the source tree below:

---

DOCUMENTATION FOR DIRECTOR:

Direktor is a continuous build system primarily meant for use by the KDE open source project. The goal was to create a system that built KDE and it's various dependencies efficiently and allowed the user to be notified and to monitor the progress of the builds.

To achieve these goals, Direktor can be configured to e-mail out the build announcements and build failures. This is augmented by the HTML data that is generated to create a dashboard for the build.

DESIGN:
Originally, Direktor was only a simple Perl script that would naively build KDE trunk and announce simple e-mailed snippets to an email address. This became fairly ungainly as time went on with the need to add in Qt4 to the build. This is when it became obvious that the build methods needed to be more abstracted away to allow any software (including KDE's base dependencies if so desired) to be built with this system.

To facilitate this, Direktor was re-factored to use description files to detail various aspects about the build and source control system per module or project. These description files are nothing more than INI files with required entries to make the system work.

A typical description file follows:

fig. 1: qt4.direktor

---


# Direktor 1.0
name = qt
repository_type = git
in_source_build = TRUE
source_repository = git://gitorious.org/+kde-developers/qt/kde-qt.git
source_directory = \$HOME/sources/qt-kde
prefix = \$HOME/installs/qt-kde
configuration_flags = -qt-gif -debug -fast -no-separate-debug-info -system-libpng -system-libjpeg -system-zlib -dbus -webkit -plugin-sql-mysql -nomake examples -nomake demos -prefix \$QTDIR
environment = QTDIR=$HOME/installs/qt-kde \
PATH=$QTDIR/bin:$PATH \
QT_PLUGIN_PATH=$HOME/installs/trunk/lib/KDE4/plugins:$QTDIR/plugins \
PKG_CONFIG_DIR=$QTDIR/lib/pkgconfig:$PKG_CONFIG_DIR \
LD_LIBRARY_PATH=$QTDIR/lib:$LD_LIBRARY_PATH \
QTEST_COLORED=1
configuration_command = configure
build_command = make
build_flags = -j2
install_command = make install


---

As can be seen, this allows the system to be very versatile with builds allowing one to build any package. The configuration_command, build_command, and install_command options allow can be any conceivable valid Linux command, even
scripts.

The fields in a direktor file are required. Lets review the fields and their format:

# Direktor 1.0
This is the description and format version of a description file as used by this application.

name:
This is an arbitrary string value. It is mainly used to present information about the module to the user.

repository_type:
This is a string value that takes one of the following values: svn, git, file, cvs, or p4. This tells Direktor which type of repository the module lives in. The file option tells Direktor that the source is not in a revision control system, but rather is a release tar.bz2, tar.gz, tgz, or zip archive.

in_source_build:
This attribute takes either a TRUE or FALSE. This boolean value tells Direktor to build the binaries from within or outside the source directory. Increasingly, many of the open source applications are built outside the source directory.

source_repository:
This is the location of the source repository or the full path location of the archive that contains the sources of the project to build.

source_directory:
This is the location where the unpacked or checked out files will be placed on the local hard disk.

prefix:
This is the prefix where the resultant binaries will be installed to.

configuration_flags:
These are the flags to pass to the script or application called to configure the source tree for compilation.

environment:
This describes the environment variables and values that must be set to allow the build of the source code to complete correctly. These are placed one per line with the continuation character being the forward slash.

configuration_command:
This tells Direktor what command or script to run in the source directory to configure the source code for compilation.

build_command:
This tells Direktor what command or script to run from the build directory to compile the source code into binary object and executable code.

build_flags:
This instructs Direktor to use the requested flags with the command to build the source tree. In many cases, this is to instruct the compiler to use non-standard library or header locations, or to do the build in parallel allowing the job to complete faster.

install_command:
This instructs Direktor what command or script to run from the build directory to install the resultant binaries produced in to the prefix earlier specified.

Note that any time an environment variable is requested that has a dollar sign in it must be escaped to properly be interpolated by the shell that is invoked by Direktor.

REGISTERING MODULES:
To register the modules, simply add the module direktor file names without the .direktor extension to the .direktor.modules_registry located in the build user's home directory. The header line declares the major series of Direktor that can read this format. An example .direktor.modules_registry follows:

fig. 2: .direktor.modules_registry:

---


# Direktor 1.0
qt4
kdesupport
kdelibs
kdepimlibs
kdebase
kdepim
kdemultimedia


---

WORK FLOW:
When Direktor starts, it clears the current environment and then sets up the global parameters defined in it's configuration stored in the home directory of the user running the build. This file is called $HOME/.direktor.cfg, and is in an extended INI format.

fig. 3: .direktor.cfg

---


# Direktor 1.0
[GENERAL]
user = somebody
group = users
registry = $HOME/.direktor.modules_registry
announce_start_via_email = 1
announce_builds_via_email = 1
announce_errors_via_email = 1
announce_completion_via_email = 1
use_dashboard = 1
; these next two are mutually exclusive!
use_watcher_to_dispatch_builds = 0
use_scheduler_to_dispatch_builds = 1
; needed for any non-generic initial repo setup that you will need. Remember,
; this gets run if the lock file $HOME/.direktor.firstrun does not exist.
first_time_startup_script = firsttimesetup.sh

[EMAIL]
from_address = someone@somewhere.net
to_address = somelist@somewhere.net
cc_addresses = bob@somewhere.net,bugs@somewhere.net
subject_prefix = MY BUILDER

[DASHBOARD]
html_upload_path = somehost.somewhere.net:/home/builder/public_html/
upload_user = someone
upload_cioher = arcfour

[WATCHER]
email_inbox = scm-monitor@somewhere.net
action_to_take = queue_all

[SCHEDULER]
queue_file = \$HOME/.direktor.schedulerrc
action_to_take = fullbuild


---

Once the global configuration is loaded, Direktor checks whether it was started in interactive or daemon mode. Once the execution mode is determined and acted on, Direktor scans the modules_registry to see which modules it needs to retrieve and build.

After it knows which targets to build, it loads the .direktor file for the given target in. Then it announces the start of the first target to the pre-defined E-Mail address and acts on the required options accordingly by setting up the additional environment needed for the compilation of the sources in this target, checking out or downloading the sources to the local disk and unpacking it if needs be to the source directory, configuring the sources, running the build command for the source target, and then installing it into a predefined prefix. Along the way, if any of the steps fail, an E-Mail announcement is sent out to notify the user building the code that the build has failed.

While this is all going on, Direktor will generate HTML pages to display the current status of the builds, including whether a build is running, the target's build completion status (success or failure), and when the build is done, what warnings were output during the compilation.

If the compile fails, the compilation log is displayed back to the last successful portion of the configuration, build, or installation.

CONTINUOUS BUILDS:
The most significant feature of Direktor is it's ability to run automatic and prescheduled builds with user specified actions. The two modes of this allow it to be used for continuous build assurance allowing projects to assure that all commits to a branch that is monitored to always build.

The first method, using the watcher system, requires that a post-commit action be set on your source control system to notify Direktor when a check-in has occurred which is delivered to a selected address configured for retrieval by a fetchmail-alike and then processed by procmail, or similar application. During the processing by procmail, the message is piped to our watcher script which then asks the scheduler to start a build as soon as the current server is idle and sends an email notification to a selected e-mail address that a commit to one of the modules has caused a new build to be queued. This very literally will on very active repositories cause the build to always be queued, so be sure that is truly the desired outcome.

The second method kicks off builds at predetermined clock times. The schedule is re-read every 15 minutes to add newly desired times that may have been added to the specified configuration file. Note that no two builds can run at the same time, so be sure that the schedule is long enough to accommodate the completion of all builds.

APP-Get Progress

I've gotten the code for the SourceManager mostly done, need a dumper function at this point and then I can utilise it in app-update. After that LWP code will be high on my list as I add the DownloadManager functions.

Code tally....

While this normally falls under the "pissing match" category, I thought this would be fun to show how much code can be shaved off of a project that is written in pure Perl versus C++ and a Shell wrapper. The old genbasedir from Apt-RPM clocks in at 1650 lines of code, not counting the includes in the C++ source files. The NEW genbasedir.pl from APP-Get clocks in at 473 lines of code. A whopping 2/3 code reduction alone. Mind you, I could have cut corners, and trimmed the code down even further, however after looking over the code that genbasedir from Apt-RPM have between the shell script and the two C++ files, there are more than a few areas that look dodgy and lack proper error checking. Tomorrow, I'll move most of the functions from genbasedir.pl to a module to make it unit testable. Stay tuned to further developments :D

Reworking genbasedir....

To make genbasedir easier to work on in the future (when we move to using some SQL engine for the data store for metadata), I reworked the functions to allow us to use an integer based pkgId instead of using the string name. This should speed up the dependency map functions in app-get considerably when doing package look ups since string comparisons aren't the fastest operations in the world in any language.

As an aside, I've gotten buy in from Jeff Johnson, the maintainer of the version of RPM at www.rpm5.org to add in some Debian inspired tags to make my work on app-get easier. :)

New APP in town :D

Just an update, I'm working on a new dep solver for RPM and the eventual general successor OPM. The utility is called APP-get, a clone of apt-get that doesn't rely on private API in the underlying package manager to work and crufty code that Apt-RPM does. To get a feel of how APP-get will retrieve the metadata, I've been reimplementing genbasedir.

At this time, I've much of the new genbasedir written. It writes out a new plain text database format for binary and source package metadata called a lst.db. Each file is a single table of fields delimited by three pipe characters in a row. With this new set of lst.db files, a person can generate the metadata using standard UNIX tools if they so wish (though I don't know why you'd want to, since genbasedir does such a nice job of it :D).

While writing this, I've been studying the dependency map algorithms used by Apt and SmartPM. Armed with the knowledge of how they determine the dependencies, I should have a fairly easy time of writing the dependency manager in APP-get.

As I've been writing the code, I'm noting a lot of places that I'll be able to clear out Debian-isms from the design, which makes the codebase and design far simpler. At the same time, I'm seeing features I'd like integrated into RPM5 to make some aspects simpler, such as package priorities.

In Apt-RPM priorities are implemented as a kludgy list file (rpmpriorities) in /etc/apt/ that is checked for a string representing the name of a given package that uses a YAML-like format that defines packages according to named priorities which are more like the Essential tag in dpkg. I'd like to change this to be INSIDE the package's header to allow distributions to mark certain packages that don't normally have a require against another as Priority: Essential, or Priority: Normal. Also if this was merged into the header of the package, it would allow rpm to gain the ability to stop people from shooting themselves in the foot with rpm -e or app-get remove.

Well, enough chatter for now. Look forward to more progress on APP-get.

I can haz workz GCC...

For the last week, I have been fighting against the GNU Compiler Collection and MJPEG Tools. Seems that there are two major bugs in GCC 4.1.2 20070110, both related to C++ code.

One is a visibility support bug, where the compiler will erroneously mark visible symbols hidden, and the other is a templating bug where redefinitions will cause an Internal Compiler Error. While the first bug would likely have caused my KDE builds to be broken, the second was blocking building kipi-plugins, since mjpegtools is a dependancy on it and does some strange redefinitions to work around a bug in the GNU Debugger.

Happily, after updating GCC to the 20070305 revision (which techinically updated us to the unofficial 4.1.3 release of GCC) mjpegtools and others seem to build as they should.

Progress

The simplicity of Qt and KDE never ceases to amaze me... With far less lines of code than if I'd been using something like WxWidgets or MSVC, I've a rich interface for an application. See the screenshots:










As can be easily seen, I've a fair amount done, and these aren't even recent screenshots. On the backend, we're likely going to be using INI and YAML for kickstart files. INI for the main portions and YAML for the partitioner though that is still under consideration.