Transcript research.cs.queensu.ca

Mining Version Histories to Guide
Software Changes
Thomas Zimmerman
Peter Weisgerber
Stephan Diehl
Andreas Zeller
“In this paper, we apply data mining to version histories: 'Programmers
who changed these functions also changed....' Just like the Amazon.com
feature helps the customer browsing along related items, our ROSE tool
guides the programmer along related changes...”
Agenda
ROSE Overview
 CVS to ROSE
 Data Analysis
 Evaluation
 Paper Critique

ROSE Overview

Aims:



Suggest and predict likely changes. Suppose a programmer has
just made a change. What else does she have to change?
Prevent errors due to incomplete changes. If a programmer wants
to commit changes, but has missed a related change, ROSE issues
a warning.
Detect coupling undetectable by program analysis. As ROSE
operates exclusively on the version history, it is able to detect
coupling between items that cannot be detected by program
analysis.
ROSE Overview (2)
CVS to ROSE




ROSE works in terms of changes in entities
 ex: changes in directories, files, classes, methods, variables
Every entity is a triple (c, i, p), where c is the syntactic category, i is
the identifier, and p is the parent entity:
 ex: (method, initDefaults(), (class, Comp, ...))
Every change is expressed using predicates:
 alter(e)
 add_to(e)
 del_from(e)
Each transaction from CVS is converted to a list of those changes
Data Analysis


ROSE aims to mine rules from those alterations:
 alter(field, fKeys[], ...) is possibly followed by:
 alter(method, initDefaults(), ...)
 alter(file, plug.properties, ...)
The probability is measured by:
 Support count. Determines the number of transactions the rule
has been derived from.
 Confidence. The relative amount of the given consequences
across all alternatives for a given antecedent.
 ex: suppose fKeys[] was altered in 11 transactions. 10 of those
also alter()'ed initDefaults() and plug.properties. 10 is the
support count, and 10/11 (or 0.909) is the confidence.
Data Analysis

Other features:



add_to() and del_from() allow an abstraction from the name of an
added entity to the name of the surrounding entity.
The notation of entities allows varying granularities for mining data.
 Fine-granular mining. For source code of C-like languages,
alter() is used for fields, functions, etc. add_to() is used for file
entities.
 Coarse-granular mining. Regardless of file type, only alter() is
used for file entities. add_to() and del_from() can be used to
capture when a file has been added or deleted
Coarse-granular rules have a higher support count and usually
return more results. However they are less precise in location, and
see limited use for guiding programmers.
Evaluation


Usage Scenarios:
 Navigation through source code. Given a change, can ROSE
point to other entities that should typically be changed too?
 Error prevention. If a programmer has changed many entities
but missed to change one, does ROSE find the missing one?
 Closure. When the transaction is finished, how often does
ROSE erroneously suggest that a change is missing in the error
prevention scenario?
Evaluation on eight large open-source projects
 ECLIPSE
 GCC
 GIMP
 JBOSS
 JEDIT
 KOFFICE
 POSTGRES
 PYTHON
Evaluation (2)

Summary:
One can have precise suggestions or many suggestions, but not
both.
 When given an initial item, ROSE makes predictions in 66 percent
of all queries. On average, the predictions of ROSE contain 33
percent of all items changed later in the same transaction, For
those queries for which ROSE makes recommendations, in 70% of
the cases, a correct location is within ROSE's topmost three
suggestions.
 In 3 percent of the queries where one item is missing, ROSE issues
a correct warning. An issued warning predicts on average 75
percent of the items that need to be considered.
 ROSE's warnings about missing items should be taken seriously:
Only 2 percent of all transactions cause a false alarm. In other
words: ROSE does not stand in the way.
 ROSE has its best predictive power for changes to existing entities.
 ROSE learns quickly: A few weeks after a project starts, ROSE
makes already useful suggestions.

Critique

Likes:
The tool was applied and accordingly evaluated to 8 projects, and
conclusions were drawn depending on their varying natures.
 It's relevant to our assignment, thus it was easy to follow.


Dislikes:
There is research value, but there is reason to be skeptical that the
“recall” of such tools will reach practical levels (for the Navigation
purposes). Intuitively, recommendations might break things if blindly
followed, regardless of if the recommendation is correct. Ie: there is
no practical value if the recommendations are incomplete, which is
more likely for complex applications where this really matters.
 I still don't know what ROSE stands for. :p
