MSR'17 presentation of our work on analyzing CI build failures.

1. An Empirical Analysis of Build Failures in the
Continuous Integration Worfklows
of Java-Based Open-Source Software
Thomas Rausch, Waldemar Hummer, Philipp Leitner*, Stefan Schulte
Distributed Systems Group
Vienna University of Technology, Austria
http://dsg.tuwien.ac.at
* Software Evolution and Architecture Lab
University of Zurich, Switzerland
http://www.ifi.uzh.ch/en/seal.html

2. 2
Continuous Integration
VCS
CI Server Build
Feedback
Logs
Vasilescu et al. (2015).
Quality and Productivity Outcomes Relating to
Continuous Integration in GitHub
“Our main finding is that continuous
integration improves the productivity of project
teams”
Kerzazi et al. (2014).
Why do Automated Builds Break? An Empirical Study
“We [...] quantified the cost of such build
breakage as more than 336.18 man-hours”

3. 3

4. 4
Related Work

5. 5
Understanding Build Failures
What types of errors cause CI build failures?
Which development practices can be
associated with CI build failures?

6. 6
Research Setting
Project Name Description
Apache Storm Distributed Computation
Butterknife Android Dependency Injection
Crate.IO Scalable SQL database
JabRef BibTeX management GUI
jcabi-github Wrapper of GitHub API
Hystrix Latency and fault tolerance library
Presto Distributed SQL query engine
Openmicroscopy Microscopy data environment
RxAndroid RxJava bindings for Android
Sponge API Minecraft plugin API
Spring Boot Java Application Framework
Square OkHttp HTTP+HTTP/2 client for Android
Square Retofit HTTP client for Android
Wordpress-Android WordPress for Android

7. 7
Data Acquisition
a
b
c
d
Topology Mapping
CI build history
Change history

8. 8
Understanding Build Failures
What types of errors cause CI build failures?
Which development practices can be
associated with CI build failures?

9. 9
Error Categorization and Quantification
 Goal
○ Categorization of errors
○ Frequency of occurrence of error types

Approach
○ Systematic exploration of ~54 000 logfiles
○ Categorization scheme based on log message patterns
[INFO] Compiling 67 source files to /home/travis/.../target/classes
[INFO] -------------------------------------------------------------
[ERROR] COMPILATION ERROR :
[INFO] -------------------------------------------------------------
[ERROR] /home/travis/.../redis/RedisAutoConfiguration.java:[143,10] cannot find symbol
[INFO] 1 error
[INFO] Compiling 67 source files to /home/travis/.../target/classes
[INFO] -------------------------------------------------------------
[ERROR] COMPILATION ERROR :
[INFO] -------------------------------------------------------------
[ERROR] /home/travis/.../redis/RedisAutoConfiguration.java:[143,10] cannot find symbol
[INFO] 1 error
[INFO] Compiling 67 source files to /home/travis/.../target/classes
[INFO] -------------------------------------------------------------
[ERROR] COMPILATION ERROR :
[INFO] -------------------------------------------------------------
[ERROR] /home/travis/.../redis/RedisAutoConfiguration.java:[143,10] cannot find symbol
[INFO] 1 error

10. 10
Error Categories
unknown Errors without a clearly identifiable cause 9
itestfailure An automated integration test failed 4
doc Documentation (e.g., JavaDoc) problem 3
license License criteria not met (missing header) 3
compatibility API incompatibility 2
androidsdk Android SDK-related error 1
buildout Error specific to Crate.IO python module 1
Label Description Occurrences
testfailure An automated test failed 12
compile Compilation error 12
git VCS interaction error 12
buildconfig Faulty build config 11
crash Build environment crash or timeout 11
dependency Dependency error 11
quality Coding-rule violation (e.g., Checkstyle) 10

11. 11
Distribution of Common Error Types
Faulty VCS
interaction
Faulty build
configuration Dependency
error
Compilation
error
Coding-rule
violation
Failing test
Crash
40%
30%
20%
10%
0%

12. 12
Distribution of Common Error Types
Apache Storm
Butterknife
Crate.IO
Hystrix
Error
testfailure
compile
git dependency crash
buildconfig quality others
Percentage
JabRef
jcabi-github
Presto
RxAndroid
SpongeAPI
Spring Boot
Square OkHttp
Square Retrofit
0% 25% 50% 75% 100%

13. 13
Understanding Build Failures
What types of errors cause CI build failures?
Which development practices can be
associated with CI build failures?

14. 14
Change Metrics
.java .txt
Changes
 Complexity
○ Churn, number of files, ...

File types
○ README.txt vs.
IntegrationTest.java
 Date and time

Author
○ Experience, commit
frequency, ...

15. 15
Process Metrics
b1
b2
b4
t
b3
a
b
e
c
f
d
g
VCS
commit graph
CI build
information
 Build History
○ Build climate
 Build Type
○ Pull request, merge, ...

Pull Request Scenarios
○ Rebase, squash, ...

16. 16
Statistical Correlation Analysis
 For each project individually
 Non-parametric correlation tests
○ Pearson’s chi-square test
○ Mann—Whitney U test

Calculate effect sizes
○ Cramér’s V
○ Rank-biserial correlation

17. 17
PassedBuild outcome Failed
Failed Passed
Previous build result
Percentageofbuilds
Findings
Build failures mostly occur consecutively.
Phases of build instability perpetuate
failures.
Build failures mostly occur consecutively.
Phases of build instability perpetuate
failures.
Build history
b
b’

18. 18
PassedBuild outcome Failed
Failed Passed
Previous build result
Percentageofbuilds
Findings
Build failures mostly occur consecutively.
Phases of build instability perpetuate
failures.
Build failures mostly occur consecutively.
Phases of build instability perpetuate
failures.
Build history
b
b’

19. 19
Findings
Build failures mostly occur consecutively.
Phases of build instability perpetuate
failures.
Build failures mostly occur consecutively.
Phases of build instability perpetuate
failures.
Build history
No evidence that either history manipula-
tion operations or parallel development
to a PR affect the PR’s build outcome.
No evidence that either history manipula-
tion operations or parallel development
to a PR affect the PR’s build outcome.
Pull request scenarios

20. 20
Findings
Even objectively harmless changes can
break builds. This indicates unwanted
flakiness of tests or the build environment.
Even objectively harmless changes can
break builds. This indicates unwanted
flakiness of tests or the build environment.
Build failures mostly occur consecutively.
Phases of build instability perpetuate
failures.
Build failures mostly occur consecutively.
Phases of build instability perpetuate
failures.
File types
Build history
 577 builds from Spring Boot
 Changelog file change only
 14% original failures
○ 52% test failures
○ 45% environment crash
○ 3% dependency error
No evidence that either history manipula-
tion operations or parallel development
to a PR affect the PR’s build outcome.
No evidence that either history manipula-
tion operations or parallel development
to a PR affect the PR’s build outcome.
Pull request scenarios

21. 21
Summary
 Categorization of error types (beyond failed/errored)
 Quantification of error type occurrence
 Statistical analysis of impact factors
 Uncovered challenges that arise when mining CI data

22. 22
Dipl.-Ing.
Thomas Rausch
Research Assistant
TU Wien
Distributed Systems Group
Argentinierstraße 8/184-1, 1040, Vienna, Austria
T: +43 1 58801 184 838
E: rausch@dsg.tuwien.ac.at
dsg.tuwien.ac.at/staff/trausch