This is a project for predicting code-change-induced configuration changes.

We use the most popular Apache Java repositories (up to Dec 2023) in our projects.

Go under ./repos, and run sh download_repos.sh to download repositories.

Go under the project root directory and run pip install requirements.txt to install necessary dependencies.

Before starting to label the data, modify configurations in conf.py to fit in your settings, e.g. OpenAI accounts. Here are the import configurations you should modify before moving on:

1. self.openai_api_key: fill in your openai api key.
2. self.label_model: fill in the expected model for labeling the data, e.g. gpt-4 or gpt-4-turbo.

Go under the project root directory and run py prepare_data.py to extract history commits and label code-change-induced configuration changes.

This step extract history commits of the projects that contain both the code and configuration file changes. We identify the code files with suffix .java, and the configuration files with suffixes .xml/.properties/.json/.yaml/.yml.

Data of the commits is saved in ./data/commit_config_related_raw/*project_name*/*commit_id*.json. In the json file, we save the new path, old path, and diff chunks of each code changes and configuration changes.

This step label whether each configuration change chunk is related to the code changes using ChatGPT. We first build prompts to ask ChatGPT whether each chunks of a configuration change is related to the changes to a code file. Following CoT techniques, we lead ChatGPT to think step by step. Here is an example the prompt format.

Here is the diff of a code change. The path of the changed file is ***. There are 2 diff chunk(s) in total.

chunk 1:

...

chunk 2:

...

Here is the diff of a configuration change. The path of the changed file is ***. There are 1 diff chunk(s) in total.

chunk1:

...

Given the code change and configuration change, which diff chunks of the configuration change are induced by the code change? Let's think step by step.

After ChatGPT returns a response, we further attach the response to a prompt which instruct ChatGPT to format the response. Here is an example the prompt format.

Respond in a json format with the chunk as the key and whether the 1 diff chunk(s) is/are induced by the code change as value (0: False, 1: True), similar as the following: {"chunk 1": 0}

We collect the responses which are in the json format, and build the labeling dataset saved in ./data/label.csv. Here is the columns of the file:

project,commit_hash,code_change_old_path,code_change_new_path,config_change_old_path,config_change_new_path,label

There is an example in ./example/label.csv.