How Do Developers React to API Deprecation?

These are the slides of our presentation at FSE 2012.

Transcript

1. How Do Developers React to API Deprecation? The Case of a Smalltalk Ecosystem Romain Robbes David Röthlisberger University of Chile Mircea Lungu Software Composition Group University of Bern FSE, 2012
2. Software Ecosystems A Natural Ecosystem
3. A Software Ecosystem A collection of software systems which are developed and co-evolve in the same environment.
4. Dependency size Activity (high) Activity (low) A Software Ecosystem A collection of software systems which are developed and co-evolve in the same environment.
5. A Software Ecosystem A Software Ecosystem A collection of software systems which are developed and co-evolve in the same environment.
6. [...] where is MOLabelShape, why was it deleted?? I use it and now is gone!!! I even had a specialization of it [...]
7. FillInTheBlank (deprecated) UIManager (recommended) Deprecating a class in the ecosystem
8. Tool Support...
9. Tool Support... Empirical Evidence First!
10. API Changes Classes Methods Providers
11. API Changes Ripple-inducing changes • addition of provider • change of provider • removal of provider Classes Methods Providers
12. API Changes Ripple-inducing changes • addition of provider • change of provider • removal of provider Classes Methods Providers Deprecation of provider = lower bounds
13. API Deprecation: Empirical Study 1. Frequency 2. Magnitude 3. Duration 4. Adaptation
14. Contributors: Projects: Classes: Commits: LOC: SqueakSource { 2.300 2.500 95.000 110.000 600.000.000 8 years
15. How to model the history of SqueakSource? Added Removed Required Provided ... Version 1 Version 2 Version i Version n ... System Version Ecosystem History History ... System History ... Artifacts History History
16. Ecco - a lighweight model Added Removed Required Provided ... Version 1 Version 2 Version i Version n ... System Version Ecosystem History History ... System History ... Artifacts History History
17. Methodology 1. Generate list of candidates (577 methods, 186 classes) 2. Filter methods removed less than 3 times 3. Manual inspection of the rest of the candidates 4. Answer RQ for all the true deprecation- based ripples
18. Manual Inspection +Retrieve changes involving deprecation +Display changes in Ripple Browser +Validate impacted projects +Track replacement methods = 180 methods + 20 classes x2
19. Results
20. Results } Frequency Magnitude Duration Adaptations
21. RQ1: Frequency of ripple effects triggered by API deprecation 14% of deprecated methods triggered ripple effects. 7% of deprecated classes triggered ripple effects.
22. RQ1: Frequency of ripple effects triggered by API deprecation 14% of deprecated methods triggered ripple effects. 7% of deprecated classes triggered ripple effects. ⤷ developers do not know their clients and program defensively
23. RQ1: Frequency of ripple effects triggered by API deprecation 14% of deprecated methods triggered ripple effects. 7% of deprecated classes triggered ripple effects. ⤷ developers do not know their clients and program defensively clients may still be unaware of deprecations ⤷
24. RQ2: Magnitude of ripple effects triggered by API deprecation reacting projects • • • • • • • • • 0 20 40 60 80 (i) reacting projects • • • • • • • • • • • 0 20 40 60 80 100 120 (ii) reacting pack
25. RQ2: Magnitude of ripple effects triggered by API deprecation reacting projects • • • • • • • • • 0 20 40 60 80 (i) reacting projects • • • • • • • • • • • 0 20 40 60 80 100 120 (ii) reacting pack ⤷the impact on the ecosystem can be very large ⤷several projects usually react
26. Viewing the magnitude from other angles Projects Packages Commits Developers 0 20 40 60 A project reacting to an API change ...
27. Viewing the magnitude from other angles Projects Packages Commits Developers ⤷may involve several developers and packages 0 20 40 60 A project reacting to an API change ...
28. Viewing the magnitude from other angles Projects Packages Commits Developers ⤷may involve several developers and packages 0 20 40 60 ⤷is often a process involving several commits A project reacting to an API change ...
29. RQ3: Duration of ripple effects triggered by API deprecation Adaptation time (in days) • • • • • • • • • • 0 200 400 600 800 (i) reaction Time 0 10 20 30 40 50 60 (ii) adaptation time For a given project, the adaptation time is the time between the first reaction and the last reaction to the API change
30. RQ3: Duration of ripple effects triggered by API deprecation Adaptation time (in days) ⤷ most projects are quick to react and adapt; but not all are • • • • • • • • • • 0 200 400 600 800 (i) reaction Time 0 10 20 30 40 50 60 (ii) adaptation time For a given project, the adaptation time is the time between the first reaction and the last reaction to the API change
31. RQ3: Duration of ripple effects triggered by API deprecation Adaptation time (in days) ⤷ most projects are quick to react and adapt; but not all are • • • • • • • • • • 0 200 400 600 800 (i) reaction Time 0 10 20 30 40 50 60 (ii) adaptation time For a given project, the adaptation time is the time between the first reaction and the last reaction to the API change ⤷ On the scale of the ecosystem, some ripple effects lasted for years.
32. RQ4: Do all projects adapt to API changes? All Active projects No counter- reactions 0 18 35 53 70 20% 40% 66%
33. RQ4: Do all projects adapt to API changes? All Active projects No counter- reactions 0 18 35 53 70 20% 40% 66% Why so few reacting projects?
34. RQ4: Do all projects adapt to API changes? All Active projects No counter- reactions 0 18 35 53 70 20% 40% 66% Dead or stagnant projects? Why so few reacting projects?
35. RQ4: Do all projects adapt to API changes? All Active projects No counter- reactions 0 18 35 53 70 20% 40% 66% Why so few reacting projects?
36. RQ4: Do all projects adapt to API changes? All Active projects No counter- reactions 0 18 35 53 70 20% 40% 66% Why so few reacting projects? Forks in the ecosystem?
37. RQ4: Do all projects adapt to API changes? All Active projects No counter- reactions 0 18 35 53 70 20% 40% 66% Why so few reacting projects?
38. RQ4: Do all projects adapt to API changes? All Active projects No counter- reactions 0 18 35 53 70 20% 40% 66%
39. RQ4: Do all projects adapt to API changes? All Active projects ⤷ common reasons for not reacting are stagnancy, forks, but also freezed dependencies and unawareness. No counter- reactions 0 18 35 53 70 20% 40% 66%
40. RQ5: Consistency of adaptations addEntity: was replaced by: add: addModel: addObject: addAll: addSibling: 0 20 40 60 80
41. RQ5: Consistency of adaptations 0 20 40 60 80 100 Probability of most likely replacement Frequency of most frequent replacement addEntity: was replaced by: add: addModel: addObject: addAll: addSibling: 0 20 40 60 80
42. RQ5: Consistency of adaptations ⤷ Many API deprecation have similar reactions 0 20 40 60 80 100 Probability of most likely replacement Frequency of most frequent replacement addEntity: was replaced by: add: addModel: addObject: addAll: addSibling: 0 20 40 60 80
43. RQ5: Consistency of adaptations ⤷ Many API deprecation have similar reactions 0 20 40 60 80 100 Probability of most likely replacement Frequency of most frequent replacement addEntity: was replaced by: add: addModel: addObject: addAll: addSibling: 0 20 40 60 80 16% of deprecations had a systematic replacement ⤷
44. RQ6: Were deprecation messages useful? foo: x self deprecated: ‘use method bar instead’
45. RQ6: Were deprecation messages useful? We categorized deprecation messages according to whether developer followed the recommendation. foo: x self deprecated: ‘use method bar instead’
46. RQ6: Were deprecation messages useful? We categorized deprecation messages according to whether developer followed the recommendation. Mostly followed: if “X” most people did X Somewhat followed: most people did Y, but others did X Not followed: most people did Y, Z ... Missing: no recommendation, vague recommendation. foo: x self deprecated: ‘use method bar instead’
47. RQ6: Were deprecation messages useful? We categorized deprecation messages according to whether developer followed the recommendation. Mostly followed: if “X” most people did X Somewhat followed: most people did Y, but others did X Not followed: most people did Y, Z ... Missing: no recommendation, vague recommendation. foo: x self deprecated: ‘use method bar instead’ self deprecated: ‘removed functionality’ self deprecated: ‘use canvas API’ self deprecated: ‘ugly method, do not call it!’
48. RQ6: Were deprecation messages useful? Mostly followed Somewhat followed Missing 0 13 25 38 50 46.8% 7.4% 40.4% Not followed 5.3% Recommendation ...
49. RQ6: Were deprecation messages useful? ⤷ half of deprecation messages are missing or not useful Mostly followed Somewhat followed Missing 0 13 25 38 50 46.8% 7.4% 40.4% Not followed 5.3% Recommendation ...
50. Threats to validity Code duplication, name clashes introduce noise We only consider explicit deprecation so far Partially manual analysis Single case study with peculiarities: – Dynamically typed language (Smalltalk) – Fork in the community
51. Related work Empirical studies: – Studies of API changes and client evolution [Dig & Johnson, 2005] – Studies of ripple effects at the system level [Yau et al., 1978; Black, 2001] Adapting to API changes: – Replaying refactorings [Henkel & Diwan 2004; Ekman & Asklund 2005; Dig et al. 2007] – Detecting systematic changes [Kim and Notkin 2009] – Adapting to evolving frameworks [Dagenais & Robillard 2008; Shäfer et al. 2008] – Filtering irrelevant changes [Holmes and Walker, 2010]
52. Conclusions and implications Added Removed Required Provided ... Version 1 Version 2 Version i Version n ... System Version Ecosystem History History ... System History ... Artifacts History History
53. Conclusions and implications ⤷ The quality of deprecation guidelines should be improved Added Removed Required Provided ... Version 1 Version 2 Version i Version n ... System Version Ecosystem History History ... System History ... Artifacts History History
54. Conclusions and implications ⤷ The quality of deprecation guidelines should be improved ⤷ Some API deprecations have a large impact Added Removed Required Provided ... Version 1 Version 2 Version i Version n ... System Version Ecosystem History History ... System History ... Artifacts History History
55. Conclusions and implications ⤷ The quality of deprecation guidelines should be improved ⤷ Some API deprecations have a large impact ⤷ Developers do not know how their software is used Added Removed Required Provided ... Version 1 Version 2 Version i Version n ... System Version Ecosystem History History ... System History ... Artifacts History History
56. Conclusions and implications ⤷ The quality of deprecation guidelines should be improved ⤷ Some API deprecations have a large impact ⤷ Developers do not know how their software is used ⤷ Reactions to API changes can be partially automated Added Removed Required Provided ... Version 1 Version 2 Version i Version n ... System Version Ecosystem History History ... System History ... Artifacts History History
57. Conclusions and implications ⤷ The quality of deprecation guidelines should be improved ⤷ Some API deprecations have a large impact ⤷ Developers do not know how their software is used ⤷ Reactions to API changes can be partially automated ⤷ Tool support is needed to help with API changes Added Removed Required Provided ... Version 1 Version 2 Version i Version n ... System Version Ecosystem History History ... System History ... Artifacts History History
58. Conclusions and implications ⤷ The quality of deprecation guidelines should be improved ⤷ Some API deprecations have a large impact ⤷ Developers do not know how their software is used ⤷ Reactions to API changes can be partially automated ⤷ Tool support is needed to help with API changes Added Removed Required Provided ... Version 1 Version 2 Version i Version n ... System Version Ecosystem History History ... System History ... Artifacts History History
59. THE END Read the details in the paper