Erupt

This is the artifact for the paper Erupt: Learned UDF Predicate Placement for Query Optimization.

Setup

conda create -n erupt python=3.9.23
conda activate erupt
python -m pip install -r requirements.txt

Run

Extract UDF features:

python udf_feature_extractor.py --db_name imdb

Training (Latency-based):

python run_train.py --db_name imdb --algo dqn
python run_train.py --db_name imdb --algo ppo

Training (Cost-based):

python run_train.py --db_name imdb --algo dqn --use_cost_model True
python run_train.py --db_name imdb --algo ppo --use_cost_model True

Test:

python run_test.py --db_name imdb --algo ppo --model_path xxxx_xxxxxx_ppo_best_policy_val.pt

Project Structure

• datasets/: datasets, workloads, UDFs.
• models/: trained models.
• cache/: cached execution plans.
• logs/: training logs.

Datasets

• IMDB: Follow instructions from IMDB; UDFs from DataManagementLab/Graceful.
• TPC-H: Download from the official TPC-H; UDFs from DataManagementLab/Graceful.
• UDFBench: Download from the official UDFBench. After loading the data, run datasets/udfbench/rename.sql to rename the columns.

Baselines

• Opt-rank: Chaudhuri and Shim, Optimization of Queries with User-Defined Predicates.
• RTOS: TsinghuaDatabaseGroup/RTOS.
• Graceful: DataManagementLab/Graceful.