On to code review lessons learned at microsoft
6 likes3,336 views
The document discusses the code review process at Microsoft, emphasizing its importance in the software development lifecycle and its potential benefits and costs. It outlines common steps in code reviews, the metrics of their usage, including over 5.5 million reviews conducted since 2010, and addresses challenges such as timeliness and the effectiveness of feedback. Additionally, the document highlights the use of mixed methods research to identify factors impacting code review quality and offers best practices for improving the process.
![Why do Code Review?
“The human eye has an almost infinite
capacity for not seeing what it does not
want to see [...] Programmers, if left to
their own devices will ignore the most
glaring errors in the output -- errors
that anyone else can see in an instant.”
Weinberg
38](https://image.slidesharecdn.com/ontocodereviewlessonslearnedatmicrosoft-160921120651/85/On-to-code-review-lessons-learned-at-microsoft-36-320.jpg)
![Better understanding. ddetter feedback.
55
“A bad reviewer tries to force their
preference on you. A good code reviewer
makes your code conform to certain
principles, but not opinion"
(Participant 4)
“Typically [a good code review] has a good
description of what the problem was, what the
solution is, and if it's a big change, it has
[documentation explaining] what it's doing and
how it's integrated with everything else.“
(Participant 4)
Better feedback.](https://image.slidesharecdn.com/ontocodereviewlessonslearnedatmicrosoft-160921120651/85/On-to-code-review-lessons-learned-at-microsoft-53-320.jpg)
On to code review lessons learned at microsoft
- 1. On to code review: Lessons learned @ Microsoft Michaela Greiler
- 2. Team: Tools for Software Engineers (TSE) • Build • Code review • Testing • Static analysis • Branch health • Software Development Practices • Empirical Software Engineering 2
- 3. Tools for Software Engineers(*) • TSE’s mission is to shorten the development cycle without impacting quality • TSE’s customers are developers (*) http://research.microsoft.com/tse Ideas Build Product Measure Data Learn
- 4. Code reviewing @Microsoft • Do code reviews make a difference? • How do we ensure an effective and efficient review process? • Does each code change need a review? • Can we automate parts of code reviews? MSR Redmond/TSE: Michaela Greiler Jacek Czerwonka Wolfram Schulte Suresh Thummalapenta Kim Herzig MSR Redmond: Christian Bird Nachi Nagappan Thomas Zimmermann MSR Cambridge: Brendan Murphy 4
- 5. Code Review is Everywhere 5
- 6. Zoom in @ Microsoft 6
- 7. Code Reviews in Engineering Workflow Code Review • Select reviewers • Review • Check-in Coding • Version control • Branch/Packages • IDE & Dev Tools Local/Buddy Build • Unit testing Pre-checkin Validation • Private test env • Functional tests • Static analysis E2E Validation • Pre-prod deploy • Scenario tests • Build drop Deploy to Production • Deployment targets Production Validation • Watchdogs • Flighting Rolling/Official Build • Buid verification • Unit testing OUTER LOOP Check-in INNER LOOP Insights 7
- 9. CodeFlow 10
- 10. Common code review steps: • Preparation of the code to be reviewed • Selection of reviewers • Notification of the selected reviewers or stakeholders • Feedback provided by reviewers • Iteration involving communication between the author and reviewer • Signoff by reviewers • Check-in of the code change to the target system 12
- 11. 13
- 12. 14
- 13. 15
- 14. 16
- 15. Code Review at Microsoft is Big • Great tool experience • Viral adoption • High user satisfaction • Natural tool consolidation • Data consolidation • 5.5M+ Code Reviews by all products since 2010 • Engineers spend more than 30 minutes a day in code reviews 17
- 16. Code Reviews Are Not Free Potential Benefits • More maintainable designs • More consistent code base • Knowledge sharing • Better awareness of changes • Additional defects found Potential Costs • Time spent by reviewers • Time spent by author addressing feedback • Time spent by change waiting in process 18
- 17. Monitoring and improvement • Several questions arise: • Is the time reviewing wise spend? • Do we find bugs in code reviews? • What is useful code review feedback? • Which issues are found during code review? • How long do engineers spend on a code review? • Which circumstances lead to more effective code reviews? • … • Drive tool improvement • New features • Tool integration • … 19
- 18. Seeking answers • Interview engineers • Perform ethnographical observations • Survey the engineers • Look at the data traces left by engineers Qualitative approach Quantitative approach 20
- 19. Mixed Method Research Is a research approach or methodology • for questions that call for real-life contextual understandings; • employing rigorous quantitative research assessing magnitude and frequency of constructs and • rigorous qualitative research exploring the meaning and understanding of constructs; DR. MARGARET-ANNE STOREY Professor of Computer Science University of Victoria All methods are inherently flawed! Generalizability Precision Realism 21
- 21. • What is codemine? What data does codemine have? 23
- 22. • What is codemine? What data does codemine have? 24
- 23. Code Review Analytics Architecture More information? See article: Lessons Learned from Building and Deploying a Code Review Analytics Platform (Christian Bird, Trevor Carnahan,Michaela Greiler) Established in June 2012 as “Garage Project” • 200GB Raw (compressed) • 500GB Curated (compressed) • > 5000 Projects • > 5M Reviews by 50K Distinct Authors • > 8M Iterations • > 8M Review Participants • > 15M Comment Threads • Running in Microsoft Azure • Won 3 internal engineering awards • Funded by Tools for Software Engineers as part of Codemine 25
- 24. CodeFlow Analytics Project Leaderboard 26
- 25. A few studies on code reviews 27
- 26. Characteristics of Useful Code Reviews: An Empirical Study at Microsoft Amiangshu Bosu, Michaela Greiler, Christian Bird 28
- 27. “Was this an impactful review, a useful comment? You know, not just a comment, but did it result in a change that wouldn’t have been there before.” - Development Team Manager 29
- 28. Study Goals •Identify factors that impact feedback usefulness • Question: What makes a code review comment useful? •Recommendations for effectiveness improvements • What factors are related to usefulness feedback? 30
- 29. Methodology: Three Phases Interview Developers Build Comment Classifier Quantitatively Investigate Factors 31
- 30. Asked author what makes code review comments useful Asked code review author to rate 20-25 comments ? Phase 1: Interview Developers 32
- 31. Useful: Pointing out defects Validation issues Info about execution environments API/tool suggestions Team conventions Designs to follow Somewhat Useful: Code formatting Typos Questions that make code more comprehensible Not Useful: Comments not in the changed code Praise Future work/Not immediately actionable Which comments are useful? 33
- 32. Positive sentiment? Is status won’t fix? Is status won’t fix? Only one comment in thread? Change within one line? Third iteration? Is status closed? Not Useful Not Useful Not Useful Not Useful Useful Useful Useful Useful 85% Recall 89% Precision Phase 2: Build Comment Classifier 34
- 33. Phase 3: Investigated Factors of review effectiveness • Classified 1.5 million reviews • Related comment usefulness to: • Characteristics of the reviewer • Characteristics of the team • Characteristics of the change under review 35
- 34. • After four reviews of a file, a developer’s knowledge reaches mastery and plateaus • It takes between six months and a year to become a high quality reviewer at MS • Quality of feedback degrades when there are more twenty files in a change More results in the paper Phase 3: Investigated Factors of review effectiveness 36
- 35. Recommendations • Include new developers for training • Include experienced developers for useful feedback • Select reviewers based on their past history with the changed files • Avoid submitting large changes for review 37
- 36. Why do Code Review? “The human eye has an almost infinite capacity for not seeing what it does not want to see [...] Programmers, if left to their own devices will ignore the most glaring errors in the output -- errors that anyone else can see in an instant.” Weinberg 38
- 37. Code Reviews Do Not Find Bugs: How the Current Code Review Best Practice Slows Us Down Jacek Czerwonka, Michaela Greiler, Jack Tilford 39
- 38. Why Code Review? 40 Reason for Code Reviewing Rank Code improvement 1 Find defects 2 Increase knowledge transfer 3 Find alternative solutions 4 Improve the development process 5 Avoid breaking builds 6 Build team awareness 7 Shared code ownership 8 Team assessment 9
- 39. What do we achieve in practice? 41
- 40. Study: Comments Classification Adapted from : M. Mäntylä and C. Lassenius. What Types of Defects Are Really Discovered in Code Reviews? IEEE Transactions Software Engineering, 35(3):430–448, 2009 By: Amiangshu Bosu (U of Alabama), Michaela Greiler (TSE), Christian Bird (Microsoft Research Redmond), Characteristics of Useful Code Reviews: An Empirical Study at Microsoft (MSR 2015) 15% of all ~50% of all 42
- 41. Study: Code Review Usefulness Adapted from : M. Mäntylä and C. Lassenius. What Types of Defects Are Really Discovered in Code Reviews? IEEE Transactions Software Engineering, 35(3):430–448, 2009 By: Amiangshu Bosu (U of Alabama), Michaela Greiler (TSE), Christian Bird (Microsoft Research Redmond), Characteristics of Useful Code Reviews: An Empirical Study at Microsoft (MSR 2015) 43
- 42. Which challenges do developer face? 44
- 43. Observations, Interviews and Survey 45
- 44. Observations, Interviews and Survey Observations & Interviews • 4 teams • 18 develops/managers • Interviews during or shortly after code review (situated insights) • Witness interactions not visible in the trace data • Highlighting cultural and social issues 46
- 45. Observations, Interviews and Survey Observations & Interviews • 4 teams • 18 develops/managers • Interviews during or shortly after code review (situated insights) • Witness interactions not visible in the trace data • Highlighting cultural and social issues Follow-up Survey • Total Respondents: 911/4,300 • Years of Experience in the Software Industry: • 2+: 87% • 6+: 70% • 10+: 40 % • Time Employed at Microsoft • 2+ years: 72% • 6+ years: 43% • 10+ years: 17% • Experience Code Reviewing • 2-5 years: 80.3% • 10+: 22% 47
- 46. Challenges 48
- 47. “Usually you write up some code and then you send it out for review, and then about a day later you ping them to remind them... and then about half a day later you go to their office and knock on their door." (Participant 7) Getting feedback in a timely manner 49
- 48. Large reviews & understanding motivation and purpose of code “It's just this big incomprehensible mess... then you can't add any value because they are just going to explain it to you and you're going to parrot back what they say." (Participant 13) 50
- 49. Challenge during Code Reviewing Overall Rank Receiving feedback in a timely manner 1 Review size 2 Managing time constraints 3 Understanding the code's purpose 4 Understanding the motivations for the change 5 Obtaining insightful feedback 6 Bikeshedding (disputing minor issues) 7 Understanding how the change was implemented 8 Maintaining code quality 9 Reaching consensus 10 Finding relevant documentation 11 Managing communication channels 12 Identifying who to talk to 13 51
- 51. Receive feedback in a timely manner 53 “Usually I try to get the person who I'm going to have review the thing to actually sit down and talk with them before I put out the code review'' (Participant 7)
- 52. Receiving Feedback in a timely manner & Managing time constraints • Author: • Aim for small, incremental changes • Allow reviewers to volunteer to perform a review • Notify potential reviewers in advance • Know when to skip a review • Run tests and analysis before • Reviewer: • Set dedicated but bounded time aside for the review • Review frequently, fewer changes but more often • Provide feedback as soon as possible 54
- 53. Better understanding. ddetter feedback. 55 “A bad reviewer tries to force their preference on you. A good code reviewer makes your code conform to certain principles, but not opinion" (Participant 4) “Typically [a good code review] has a good description of what the problem was, what the solution is, and if it's a big change, it has [documentation explaining] what it's doing and how it's integrated with everything else.“ (Participant 4) Better feedback.
- 54. Better understanding. Better feedback. • Author: • Aim for small, incremental changes • Cluster related changes and submit changes with context • Document the motivation • Select reviewers with the right expertise • Show gratitude & carefully consider feedback • Reviewer: • Use richer communication channels, e.g. face-to-face meetings • Use tools that provide traceability for non-contentious or sensitive issues • Give constructive and respectful feedback • Justify and explain the reasons for rejecting a change • Focus on core issues first 56
- 55. As an organization, establish a review policy that promotes: • Build a positive review culture that sets the tone for feedback • Ensure time spent reviewing is “counted" and “expected" • Watch for negative impacts of employee assessment or incentives • Ensure appropriate tools are used • Make sure tools integrate well • Ensure sufficient training (e.g. junior working alongside senior) • Develop, reflect on, and revise code reviewing policies and checklists. • Resolve bottlenecks • Identifying inappropriate or aggressive communication 57
- 56. Tool improvements • Enforce a reviewing workflow • Support finding right experts • Show test coverage of code under review • Automatic clustering of changes into groups of related entities • Automate feedback • Traceability support • Dashboards for metrics and pending reviews • Management for notifications • Integration with other communication tools 58
- 57. Tool integration • Trigger static analysis / inspect analysis output • Trigger testing / inspect test output • Overlay code with test coverage • Trigger builds / inspect build output • Automatic reviewer selection • Check-in code 59
- 58. Code review video (3:40) 60
- 59. Michaela Greiler @mgreiler www.michaelagreiler.com http://research.microsoft.com/tse 61 Challenge during Code Reviewing Rank Receiving feedback in a timely manner 1 Review size 2 Managing time constraints 3 Understanding the code's purpose 4