QPPNet in PyTorch

The README below was written for the release tagged rabbit721.

This code contains a sample implementation for Plan-Structured Deep Neural Network Models for Query Performance Prediction presented at VLDB 2019, and the code for training/testing on

• TPC-H queries generated using https://github.com/gregrahn/tpch-kit.git and benchmarked with Postgres

  The TPC-H data are generated with ./dbgen -s <scale factor>.

  The TPC-H queries are generated using ./qgen <query_id> -r <seed number> where varying seed random number is used to generate different queries from a template.

  Tables in PostgresSQL are created using dss.ddl and then have indexes created with tpch-postgres-index-ddl.sql.

  The query plan structure and query performance metrics are retrieved as json objects by running the generated queries above with the explain (analyze, format JSON, verbose) statement prepended.

• TPC-H queries generated using https://github.com/gregrahn/tpch-kit.git and benchmarked with NoisePage

• TPC-C queries and smallbank dataset generated using OLTP and benchmarked with NoisePage

Prerequisites

• Linux or macOS
• Python 3

Getting Started

Installation

• Cloning the repo:

  git clone https://github.com/rabbit721/QPPNet.git
  cd QPPNet
  

• Install the required python packages:

  • For pip: pip install -r requirements.txt
  • For conda: conda env create -f environment.yml

Getting the Datasets

• TPC-H benchmarked with PostgresSQL

  # Under directory dataset/postgres_tpch_dataset
  wget http://www.andrew.cmu.edu/user/jiejiao/data/qpp/postgres/tpch/psqltpch0p1g.zip && unzip psqltpch0p1g.zip
  wget http://www.andrew.cmu.edu/user/jiejiao/data/qpp/postgres/tpch/psqltpch1g.zip && unzip psqltpch1g.zip
  wget http://www.andrew.cmu.edu/user/jiejiao/data/qpp/postgres/tpch/psqltpch10g.zip && unzip psqltpch10g.zip
  

• TPC-H benchmarked with NoisePage:

  Data files already located under directory dataset/terrier_tpch_dataset as execution_0p1G.csv, execution_1G.csv, and execution_10G.csv

• TPC-C and smallbank benchmarked with NoisePage:

  Data files already located under directory dataset/oltp_dataset as tpcc_pipeline.csv and sb_pipeline.csv

Examples for Training a model

• On TPC-H dataset generated using https://github.com/gregrahn/tpch-kit.git with SF=1 and benchmarked with PostgresSQL

  python3 main.py --dataset PSQLTPCH -s 0 -t 250000 --batch_size 128 -epoch_freq 1000 --lr 2e-3 --step_size 1000 --SGD --scheduler --data_dir ./dataset/postgres_tpch_dataset/tpch1g/900-exp_res_by_temp/  
  

• On TPC-H dataset generated using https://github.com/gregrahn/tpch-kit.git with SF=1 and benchmarked with NoisePage

  python3 main.py --dataset TerrierTPCH -s 0 -t 250000 --batch_size 512 -epoch_freq 1000 --lr 1e-3 --step_size 1000 --SGD --scheduler --data_dir ./dataset/terrier_dataset/execution_1G.csv
  

• On TPC-C dataset generated using OLTP with SF=1 and benchmarked with NoisePage

  python3 main.py --dataset OLTP -s 0 -t 250000 --batch_size 512 -epoch_freq 1000 --lr 5e-3 --step_size 1000 --SGD --scheduler --data_dir ./dataset/oltp_dataset/tpcc_pipeline.csv
  

Using a pre-trained model

• Getting a model trained for 4000 epochs on TPC-H SF=1 dataset benchmarked with PostgresSQL:

  wget http://www.andrew.cmu.edu/user/jiejiao/data/qpp/trained_models/psqltpch_epoch4000.zip
  

• Getting a model trained for 20000 epochs on TPC-H SF=1 dataset benchmarked with NoisePage:

  wget http://www.andrew.cmu.edu/user/jiejiao/data/qpp/trained_models/terriertpch_epoch20000.zip
  

• Getting a model trained for 10000 epochs on TPC-C dataset generated with OLTP and benchmarked with NoisePage:

  wget http://www.andrew.cmu.edu/user/jiejiao/data/qpp/trained_models/tpcc_epoch10000.zip
  

Examples for Testing a trained model

When testing, please make sure that trained models are saved in ./saved_model and the mean and range values of the train data are provided with the flag '--mean_range_dict' in order to normalize the input.

The '-s' flag is interpreted as an integer and is used to specify the epoch number of the saved model to be tested.

• Testing a model trained for 4000 epochs on TPC-H SF=1 dataset benchmarked with PostgresSQL on TPC-H SF=10 dataset benchmarked with PostgresSQL.

  python3 main.py --test_time --dataset PSQLTPCH -s 4000 --mean_range_dict mean_range_dict.pickle --data_dir ./dataset/postgres_tpch_dataset/tpch10G/900-exp_res_by_temp/
  

• Testing a model trained for 10000 epochs on the TPC-C benchmark on the smallbank dataset. Please make sure that models are saved in ./saved_model

  python3 main.py --test_time --dataset OLTP -s 10000 --data_dir ./dataset/oltp_dataset/sb_pipeline.csv