Automatically Generating Dockerfiles via Deep Learning: Challenges and Promises
0 likes105 views
This document discusses the challenges and advancements in automatically generating Dockerfiles using deep learning techniques, particularly through the T5 model. It highlights the time-consuming nature of writing Dockerfiles and the limitations of current AI-based tools, while presenting a high-level specification approach for better code generation. The study finds that the T5 model slightly outperforms traditional methods but still requires manual adjustments and faces challenges related to the training data and procedures.
![High-Level specification
HLS
Operating System: “alpine”
Package Manager: “apk”
Package Requirements: [“python3”]
Download from Sources: FALSE
ENV variables: FALSE
Build ARGs: FALSE
LABEL: TRUE
EXPOSE for ports: TRUE
CMD: TRUE
ENTRYPOINT: FALSE
High-Level Specification](https://image.slidesharecdn.com/icssp23dockergen-230514030640-d589eec5/85/Automatically-Generating-Dockerfiles-via-Deep-Learning-Challenges-and-Promises-22-320.jpg)
![High-Level specification
HLS
Operating System: “alpine”
Package Manager: “apk”
Package Requirements: [“python3”]
Download from Sources: FALSE
ENV variables: FALSE
Build ARGs: FALSE
LABEL: TRUE
EXPOSE for ports: TRUE
CMD: TRUE
ENTRYPOINT: FALSE
High-Level specification
High-Level Specification](https://image.slidesharecdn.com/icssp23dockergen-230514030640-d589eec5/85/Automatically-Generating-Dockerfiles-via-Deep-Learning-Challenges-and-Promises-23-320.jpg)
![High-Level specification
HLS
Operating System: “alpine”
Package Manager: “apk”
Package Requirements: [“python3”]
Download from Sources: FALSE
ENV variables: FALSE
Build ARGs: FALSE
LABEL: TRUE
EXPOSE for ports: TRUE
CMD: TRUE
ENTRYPOINT: FALSE
High-Level specification
High-Level Specification](https://image.slidesharecdn.com/icssp23dockergen-230514030640-d589eec5/85/Automatically-Generating-Dockerfiles-via-Deep-Learning-Challenges-and-Promises-24-320.jpg)
Automatically Generating Dockerfiles via Deep Learning: Challenges and Promises
- 1. Automatically Generating Dockerfiles via Deep-Learning: Challenges and Promises Giovanni Rosa, Antonio Mastropaolo, Simone Scalabrino, Gabriele Bavota and Rocco Oliveto 17th International Conference on Software and System Processes 14-15 May 2023, MCEC (Melbourne) @giovannipink University of Molise, Italy
- 2. Containerization, how it works? Build Ship Run Software Containers
- 3. 85% of organizations will adopt containers by 2025
- 4. #1 Most-Wanted and #1 Most Loved tool 85% of organizations will adopt containers by 2025
- 5. Dockerfile Image Container run build Docker in a Nutshell
- 6. Dockerfile Image Container run build Docker in a Nutshell
- 7. Dockerfile Container run build Docker in a Nutshell Image
- 8. Dockerfile Image Container run build Docker in a Nutshell
- 10. Time-consuming activity Reis et. al 2021 Henkel et. al 2020
- 11. Time-consuming activity Technical knowledge required Reis et. al 2021 Henkel et. al 2020
- 12. Time-consuming activity Technical knowledge required Limited supporting tools Reis et. al 2021 Henkel et. al 2020
- 13. Hanayama et. al 2020 AI-based Code Assistants …limited to dependencies and code completion Horton et. al 2019 Ye et. al 2021 Hanayama et. al 2020
- 14. Horton et. al 2019 Ye et. al 2021 Hanayama et. al 2020 AI-based Code Assistants …limited to dependencies and code completion
- 15. The automated generation from scratch is still an open challenge
- 17. Not Bad! but… build error L L Prompt: Can you provide a Dockerfile using Debian 9 and installing Golang 1.9.4 via sources?
- 18. What about state-of-the-art DL models for code-related tasks? T5
- 20. T5 Model Generated Dockerfile T5 prediction Dockerfile Generation via T5 input Natural language requirements
- 21. How to represent software requirements for a Dockerfile? Natural Language: Too Broad!
- 22. High-Level specification HLS Operating System: “alpine” Package Manager: “apk” Package Requirements: [“python3”] Download from Sources: FALSE ENV variables: FALSE Build ARGs: FALSE LABEL: TRUE EXPOSE for ports: TRUE CMD: TRUE ENTRYPOINT: FALSE High-Level Specification
- 23. High-Level specification HLS Operating System: “alpine” Package Manager: “apk” Package Requirements: [“python3”] Download from Sources: FALSE ENV variables: FALSE Build ARGs: FALSE LABEL: TRUE EXPOSE for ports: TRUE CMD: TRUE ENTRYPOINT: FALSE High-Level specification High-Level Specification
- 24. High-Level specification HLS Operating System: “alpine” Package Manager: “apk” Package Requirements: [“python3”] Download from Sources: FALSE ENV variables: FALSE Build ARGs: FALSE LABEL: TRUE EXPOSE for ports: TRUE CMD: TRUE ENTRYPOINT: FALSE High-Level specification High-Level Specification
- 25. OS: “alpine3.17” Pkg. Manager: “apk” Pkg. Requirements: “python3” High-Level specification High-Level specification High-Level Specification
- 26. 12 software developers +2 years > 50% agrees with the requirements specification High-Level specification Asking Developers’ Opinion
- 27. 9.4M unique Dockerfiles from 2013 to 2020 Eng et. al 2021 High-Level specification Dockerfile Dataset
- 31. 100k Model Tuning 11k Test Resulting Dataset
- 33. 3pre-training settings Dockerfile & English Dockerfile Only English Only T5 model construction
- 34. T5 model construction Dockerfile & English Dockerfile Only English Only Fine Tuning
- 37. Elasticsearch HLS ES node bool query IR-Baseline 1
- 40. Adherence to the input High-Level Specification RQ1 Metric: Field-by-field match +1 +1 +1 +1 +1 +1 +1 +1 +1 +1 Input HLS Generated/Retrieved HLS VS
- 41. ES Results for RQ1 OS ~1.00 0.92 0.88 Pkg. Manager 0.98 1.00 1.00 Pkg. Requirements 0.87 0.88 0.76 Download from sources 0.82 0.84 0.52 ENV variables 0.89 0.81 0.17 Build ARGs 0.99 0.88 0.17 LABEL ~1.00 0.87 0.37 EXPOSE for ports 0.80 0.83 0.45 CMD 0.74 0.83 0.26 ENTRYPOINT 0.84 0.85 0.45 T5
- 42. Structural similarity between Dockerfiles RQ2 Metric: AST edit distance Input Dockerfile Parsed AST (Henkel et. al 2020)
- 43. RQ2 Results for RQ2 The lower is better Structural similarity between Dockerfiles ES T5
- 44. RQ3 Metric: Percentage of matching layers build Input Dockerfile build Gen./Retr. Dockerfile Docker Image A Docker Image B Matching SHAs Similarity between Docker images
- 45. Similarity between Docker images RQ3 Results for RQ3 ES T5
- 46. Summary T5 achieves slightly better results than IR ...
- 47. Summary T5 achieves slightly better results than IR ... Generated Dockerfiles require manual adjustments More resource-consuming compared to IR
- 48. What we have learned?
- 49. Challenge #1 Not enough training instances
- 50. Challenge #1 Not enough training instances Data augmentation
- 51. Challenge #2 A different training procedure must be used
- 52. Challenge #2 Different stopping criterion A different training procedure must be used Dockerfile abstractions