Code Maat is a command line tool used to mine and analyze data from version-control systems (VCS).
.
Code Maat was developed to accompany the discussions in my book Your Code as a Crime Scene. In case you're interested in repository mining, check-out the book here.
To understand large-scale software systems we need to look at their evolution. The history of our system provides us with data we cannot derive from a single snapshot of the source code. Instead VCS data blends technical, social and organizational information along a temporal axis that let us map out our interaction patterns in the code. Analyzing these patterns gives us early warnings on potential design issues and development bottlenecks, as well as suggesting new modularities based on actual interactions with the code. Addressing these issues saves costs, simplifies maintenance and let us evolve our systems in the direction of how we actually work with the code.
Maat was a goddess in ancient Egyptian myth. She was the one who gave us order out of the initial chaos in the universe. Code Maat hopes to continue the work of Maat, albeit on a smaller basis, by highlighting code with chaotic development practices and suggest the directions of future refactorings to bring order to it. Further, maat was used in ancient Egypt as a concept of truth. And metrics never lie (except when they do).
Copyright © 2013-2015 Adam Tornhill
Distributed under the GNU General Public License v3.0.
Currently, I'm not hosting any pre-built binaries. Code Maat is written in Clojure. To build it from source, use leiningen:
lein uberjar
The command above will create a standalone jar containing all the dependencies.
Code Maat operates on log files from version-control systems. The supported version-control systems are git, Mercurial (hg), svn, and Perforce (p4). The log files are generated by using the version-control systems themselves as described in the following sections.
To analyze our VCS data we need to define a temporal period of interest. Over time, many design issues do get fixed and we don't want old data to interfere with our current analysis of the code. To limit the data Code Maat will consider, use one of the following flags depending on your version-control system:
- git: Use the
--after=<date>to specify the last date of interest. The<date>is given asYYYY-MM-DD. - hg: Ue the
--dateswith to specify the last date of interest. The value is given as">YYYY-MM-DD". - svn: Use the
-roption to specify a range of interest, for example-r {20130820}:HEAD. - p4: Use the
-moption to specify the last specified number of changelists, for example-m 1000.
svn log -v --xml > logfile.log -r {YYYYmmDD}:HEAD
git log --pretty=format:'[%h] %aN %ad %s' --date=short --numstat --after=YYYY-MM-DD
Note that there's a second supported Git format as well, imaginatively named git2. This format is more tolerant and faster to parse, so please prefer it over the plain git format described above:
git log --all -M -C --numstat --date=short --pretty=format:'--%h--%ad--%an' --no-renames
hg log --template "rev: {rev} author: {author} date: {date|shortdate} files:\n{files %'{file}\n'}\n" --date ">YYYY-MM-DD"
p4 changes -s submitted -m 5000 //depot/project/... | cut -d ' ' -f 2 | xargs -I commitid -n1 sh -c 'p4 describe -s commitid | grep -v "^\s*$" && echo ""'
You can run Code Maat directly from leiningen:
lein run -l logfile.log -c <vcs>
If you've built a standalone jar (lein uberjar), run it with a simple java invocation:
java -jar code-maat-0.8.6.jar -l logfile.log -c <vcs>
When invoked with -h, Code Maat prints its usage:
adam$ java -jar code-maat-0.9.0.jar
This is Code Maat, a program used to collect statistics from a VCS.
Version: 0.9.0-SNAPSHOT
Usage: program-name -l log-file [options]
Options:
-l, --log LOG Log file with input data
-c, --version-control VCS Input vcs module type: supports svn, git, git2, hg, or p4
-a, --analysis ANALYSIS authors The analysis to run (abs-churn, age, author-churn, authors, communication, coupling, entity-churn, entity-effort, entity-ownership, fragmentation, identity, main-dev, main-dev-by-revs, messages, refactoring-main-dev, revisions, soc, summary)
--input-encoding INPUT-ENCODING Specify an encoding other than UTF-8 for the log file
-r, --rows ROWS Max rows in output
-g, --group GROUP A file with a pre-defined set of layers. The data will be aggregated according to the group of layers.
-n, --min-revs MIN-REVS 5 Minimum number of revisions to include an entity in the analysis
-m, --min-shared-revs MIN-SHARED-REVS 5 Minimum number of shared revisions to include an entity in the analysis
-i, --min-coupling MIN-COUPLING 30 Minimum degree of coupling (in percentage) to consider
-x, --max-coupling MAX-COUPLING 100 Maximum degree of coupling (in percentage) to consider
-s, --max-changeset-size MAX-CHANGESET-SIZE 30 Maximum number of modules in a change set if it shall be included in a coupling analysis
-e, --expression-to-match MATCH-EXPRESSION A regex to match against commit messages. Used with -messages analyses
-t, --temporal-period TEMPORAL-PERIOD Instructs Code Maat to consider all commits during the same day as a single, logical commit
-d, --age-time-now AGE-TIME_NOW Specify a date as YYYY-MM-dd that counts as time zero when doing a code age analysis
-h, --help
By default, Code Maat expects your log files to be UTF-8. If you use another encoding, override the default with --input-encoding, for example --input-encoding UTF-16BE.
When starting out, I find it useful to get an overview of the mined data. With the summary analysis, Code Maat produces such an overview:
java -jar code-maat-0.9.0.jar -l logfile.log -c git -a summary
The resulting output is on csv format:
statistic, value
number-of-commits, 919
number-of-entities, 730
number-of-entities-changed, 3397
number-of-authors, 79
If you use the second Git format, just specify git2 instead:
java -jar code-maat-0.9.0.jar -l logfile2.log -c git2 -a summary
By default, Code Maat runs an analysis on the number of authors per module. The authors analysis is based on the idea that the more developers working on a module, the larger the communication challenges. The analysis is invoked with the following command:
java -jar code-maat-0.9.0.jar -l logfile.log -c git
The resulting output is on CSV format:
entity, n-authors, n-revs
InfoUtils.java, 12, 60
BarChart.java, 7, 30
Page.java, 4, 27
...
In example above, the first column gives us the name of module, the second the total number of distinct authors that have made commits on that module, and the third column gives us the total number of revisions of the module. Taken together, these metrics serve as predictors of defects and quality issues.
Logical coupling refers to modules that tend to change together. Modules that are logically coupled have a hidden, implicit dependency between them such that a change to one of them leads to a predictable change in the coupled module. To analyze the logical coupling in a system, invoke Code Maat with the following arguments:
java -jar code-maat-0.9.0.jar -l logfile.log -c git -a coupling
The resulting output is on CSV format:
entity, coupled, degree, average-revs
InfoUtils.java, Page.java, 78, 44
InfoUtils.java, BarChart.java, 62, 45
...
In the example above, the first column (entity) gives us the name of the module, the second (coupled) gives us the name of a logically coupled module, the third column (degree) gives us the coupling as a percentage (0-100), and finally average-revs gives us the average number of revisions of the two modules. To interpret the data, consider the InfoUtils.java module in the example output above. The coupling tells us that each time it's modified, it's a 78% risk/chance that we'll have to change our Page.java module too. Since there's probably no reason they should change together, the analysis points to a part of the code worth investigating as a potential target for a future refactoring.
The change frequency of code is a factor that should (but rarely do) drive the evolution of a software architecture. In general, you want to stabilize as much code as possible. A failure to stabilize means that you need to maintain a working knowledge of those parts of the code for the life-time of the system.
One way to measure the stability of a software architecture is by a code age analysis:
java -jar code-maat-0.9.0.jar -l logfile.log -c git -a age
The age analysis grades each module based on the date of last change. The measurement unit is age in months. Here's how the result may look:
entity,age-months
src/code_maat/app/app.clj,2
project.clj,4
src/code_maat/parsers/perforce.clj,5
...
By default, Code Maat uses the current date as starting point for a code age analysis. You specify a different start time with the command line argument --age-time-now.
By using the techniques from Your Code as a Crime Scene we visualize the system with each module marked-up by its age (the more red, the more recent changes to the code):
.
I've chosen to keep the visualizations separate from the analysis engine itself. It just gives you more flexibility. After all, CSV gives you a nice format to transform and experiment with.
I present a whole suite of different visualization techniques and options in Your Code as a Crime Scene, so do check out the book if you want to dive deeper. You can also look at some of the tools I've open sourced such as Metrics Tree Map:
.
An alternative is to save the generated CSV to a file and import it into a spreadsheet program such as OpenOffice or Excel. That allows us to generate charts such as the ones below:
.
Code churn is related to post-release defects. Modules with higher churn tend to have more defects. There are several different aspects of code churn. I intend to support several of them in Code Maat.
The absolute code churn numbers are calculated with the -a abs-churn option. Note that the option is only available for git. The analysis will output a CSV table with the churn accumulated per date:
date, added, deleted
2013-08-09, 259, 20
2013-08-19, 146, 77
2013-08-21, 5, 6
2013-08-20, 773, 121
2013-08-30, 349, 185
...
Visualizing the result allows us to spot general trends over time:
.
The idea behind this analysis is to get an idea of the overal contributions by each individual. The analysis is invoked with the -a author-churn option. The result will be given as CSV:
author, added, deleted
Adam Tornhill, 13826, 1670
Some One Else, 123, 80
Mr Petersen, 3, 3
...
And, of course, you wouldn't use this data for any performance evaluation; it wouldn't serve well (in case anything should be rewarded it would be a net deletion of code - there's too much of it in the world).
The pre-release churn of a module is a good predictor of its number of post-release defects. Such an analysis is supported in Code Maat by the -a entity-churn option.
Note: Some research suggests that relative churn measures are better, while others don't find any significant differences. The metrics calculated by Code Maat are absolute for now because it's easier to calculate. I'm likely to include support for relative churn too.
Once we have mined the organizational metrics descibed above, we may find we have multiple developers working on the same modules. How is their effort distributed? Does a particular module have a major developer or is everyone contributing a small piece? Let's find out by running the -a entity-ownership analysis. This analysis gives us the following output:
entity, author, added, deleted
analysis/authors.clj, apt, 164, 98
analysis/authors.clj, qew, 81, 10
analysis/authors.clj, jt, 42, 32
analysis/entities.clj, apt, 72, 24
...
Another ownership view is to consider the effort spent by individual authors on the different entities in the system. This analysis is run by the -a entity-effort option. The analysis gives us the following table:
entity, author, author-revs, total-revs
analysis/authors.clj, apt, 5, 10
analysis/authors.clj, qew, 3, 10
analysis/authors.clj, jt, 1, 10
analysis/authors.clj, apt, 1, 10
...
This information may be a useful guide to find the right author to discuss functionality and potential refactoringswith. Just note that the ownership metrics are sensitive to the same biases as the churn metrics; they're both heuristics and no absolute truths.
Sometimes we'd like to find patterns that manifests themselves over multiple commits. Code Maat provides the --temporal-period switch that let you consider all commits within a day as a logical change. Just provide the switch and add a digit - in the future that digit may even mean something; Right now the aggregation is limited to commits within a single day.
Code Maat supports an identity analysis. By using this switch, Code Maat will output the intermediate parse result of the raw VCS file. This can be useful either as a debug aid or as input to other tools.
Code Maat uses the Incanter library. By default, Incanter will create an awt frame. You can surpress the frame by providing the following option to your java command: -Djava.awt.headless=true.
Code Maat is quite memory hungry, particularly when working with larger change sets. Thus, I recommend specifying a larger heap size than the JVM defaults: -Xmx4g.
Note that when running Code Maat through leiningen, those options are already configured in the project.clj file.
The current version of Code Maat processes all its content in memory. Thus, it may not scale to large input files (however, it depends a lot on the combination of parser and analysis). The recommendation is to limit the input by specifying a sensible start date (as discussed initially, you want to do that anyway to avoid confounds in the analysis).
In future versions of Code Maat I plan to add more analysis methods such as code churn and developer patterns. I also plan on direct visualization support and a database backed analysis to allow processing of larger log files. Further, I plan to add a worked example. That example will be a case study of some well-known open source code. Until then, I hope you find Code Maat useful in its initial shape.