This is the source code of paper DistJoin: A Decoupled Join Cardinality Estimator based on Adaptive Neural Predicate Modulation

Env setup

1. Install python3.12
2. Install required packages in requirements.txt python install -r requirements.txt
3. Install our sampler package that response for generating training data dynamically during training python install ./MySampler/setup.py install

Prepare Dataset

1. Put the JOB datasets into ./datasets/job, all csv table should have headers
2. You can update the true cards by first removing all {wokload}.pkl file in ./queries and run the ./queries/ConvertMSCNTestWorkload.py, which will automatically calculates true cards and convert the test workloads to MSCN's format
3. Use ./queries/GetJoinWithoutPredicatesCard.py to pre-calculates the cardinality of queries' join schemas if needed

Setup experiments

1. Use ./Configs/IMDB/IMDB.yaml to set experiments, or you can use the default one to perform our experiments in the paper

Train DistJoin

1. Run python train.py
2. Copy the exp mark in the output for latter testing, which is a timestamp

Test DistJoin

1. Run python eval-IMDB-all.py --config=IMDB --no_wandb and enter the exp mark to evaluate the workloads configurated in the IMDB.yaml file, it will cover all five join conditions on that workload
2. Check the results in the output and the ./results/DistJoin