Artifacts accompanying LasRCA, a framework for one-shot root cause analysis in cloud-native systems that leverages the collaboration of the LLM and the small classifier.
cd ./code
pip install -r requirements.txtThe model and temp data needs to be downloaded from https://drive.google.com/drive/folders/1GK2GrJ6zxMI_cOCm91sWt38d_pUIQb0q?usp=sharing.
Then the model and temp_data folders need be placed following the structure of the working folder:
.
├── README.md
├── code
│ ├── data_filter preprocess data
│ ├── llm generate LLM requests, save responses and process to fault labels
│ ├── pretrain pretrain TS2Vec models for services, pods, and nodes (If there is no need to conduct experiments in motivation, it can be ignored)
│ ├── shared_util some basic util functions like load/save pickle files, set deterministic seeds, etc.
│ ├── tasks main experiments
│ │ ├── models components of the small classifier
│ │ ├── base.py base classes for RCA
│ │ ├── base_fault_type_classification.py semi-supervised learning strategy and some other strategies for RCA
│ │ ├── dataset.py load dataset
│ │ ├── eval.py evaluation metrics
│ │ ├── fine_tuning_fault_type_classification.py pretrain and fine-tuning strategy for RCA
│ │ ├── las_rca.py main experiments of LasRCA
│ │ └── util some util functions for different strategies
│ ├── pretrain.sh shell for pretrain TS2Vec models
│ ├── reproduce.sh shell for reproducibility
│ ├── reproduce.log the execution log for reproducibility
│ └── requirements.txt
├── model saved model data for reproduction (place the extracted ``model`` folder here)
└── temp_data saved temp data for reproduction (place the extracted ``temp_data`` folder here)requirements.txt.model.zip and temp_data.zip. Then move the folders model and temp_data to the above correct situation.Note: Although the path is configurable in our code, we strongly recommend placing the code on a path consistent with our experiment (/workspace/project/working/2024/), as this may avoid additional path parameter settings.
cd ./code
bash reproduce.shSince the raw data is too big, we list the link here, help for downloading. The link is https://competition.aiops-challenge.com/home/competition/1496398526429724760. (Sometimes the page may be crashed, please visit https://www.bizseer.com/index.php?m=content&c=index&a=show&catid=25&id=83 for simple introduction).
Note: Although the path is configurable in our code, we strongly recommend placing the code on a path consistent with our experiment (/workspace/project/working/2024/), as this may avoid additional path parameter settings.
First put the raw data downloaded from the above link into the path described in code/data_filter/CCF_AIOps_challenge_2022/config/dataset_config.py. (The default path is /workspace/dataset/2022_CCF_AIOps_challenge/). The folder structure is as follows:
/workspace/dataset/2022_CCF_AIOps_challenge
├── test_data
├── training_data_normal
└── training_data_with_faultsThen run the following Python scripts to preprocess raw data.
cd ./code
python ./data_filter/CCF_AIOps_challenge_2022/dao/ground_truth_dao.py
python ./data_filter/CCF_AIOps_challenge_2022/dao/topology_dao.py
python ./data_filter/CCF_AIOps_challenge_2022/dao/metric_dao.py
python ./data_filter/CCF_AIOps_challenge_2022/dao/log_dao.py
python ./data_filter/CCF_AIOps_challenge_2022/dao/trace_dao.py
python ./data_filter/CCF_AIOps_challenge_2022/service/time_interval_label_generator.py
python ./data_filter/CCF_AIOps_challenge_2022/service/ent_edge_index_generator.py
python ./data_filter/CCF_AIOps_challenge_2022/service/metric_generator.py
python ./data_filter/CCF_AIOps_challenge_2022/service/log_generator.py
python ./data_filter/CCF_AIOps_challenge_2022/service/trace_generator.py
python ./data_filter/CCF_AIOps_challenge_2022/service/dataset_generator.pyThen you will find a new folder named temp_data that is side by side with the code.
If you want to replicate the pretraining and fine-tuning experiments mentioned in the paper, you need to first pretrain the embedding. Run the following command:
cd ./code
bash pretrain.shPretrain on TS2Vec../code/tasks/base_fault_type_classification.py, and comment the 316-338 lines in ./code/tasks/base_fault_type_classification.py../code/tasks/fine_tuning_fault_type_classification.py, and comment the 303-324 lines in ./code/tasks/fine_tuning_fault_type_classification.py.python ./tasks/fine_tuning_fault_type_classification.py (for Fine-tuning)
python ./tasks/base_fault_type_classification.py (for Semi-supervised Learning)./code/tasks/las_rca.py, and comment the 275-297 lines in ./code/tasks/las_rca.py.k in the 211 line to 0.python ./tasks/las_rca.py.Before conducting LLM Fault Labeling, it is necessary to consider the correct settings of the following parameters:
k in the 7 line of code/llm/common.py: The iteration of LLM fault labeling. We use the value 0 as default. Therefore, LLM fault labeling is only executed once by default.api_key in the 8 line of code/llm/common.py: Due to the high cost of GPT's API, we cannot disclose our api_key. You can switch api_key to your own account.Then run the following scripts:
python ./llm/first_stage.py
python ./llm/second_stage.py
python ./llm/third_stage.py
python ./llm/summarization_stage.py
python ./llm/update_dataset.py./code/tasks/las_rca.py, and comment the 275-297 lines in ./code/tasks/las_rca.py.k in the 211 line to 1, and the parameter k in the 212 line to "GPT4".python ./tasks/las_rca.py.