DeepDB: Learn from Data, not from Queries!
DeepDB is a data-driven learned database component achieving state-of-the-art-performance in cardinality estimation and approximate query processing (AQP). This is the implementation described in
Benjamin Hilprecht, Andreas Schmidt, Moritz Kulessa, Alejandro Molina, Kristian Kersting, Carsten Binnig: "DeepDB: Learn from Data, not from Queries!", VLDB'2020. [PDF]
Setup
Tested with python3.7 and python3.8
git clone https://github.com/DataManagementLab/deepdb-public.git
cd deepdb-public
sudo apt install -y libpq-dev gcc python3-dev
python3 -m venv venv
source venv/bin/activate
pip3 install -r requirements.txt
For python3.8: Sometimes spflow fails, in this case remove spflow from requirements.txt, install them and run
pip3 install spflow --no-deps
Reproduce Experiments
Cardinality Estimation
Download the Job dataset. Generate hdf files from csvs.
python3 maqp.py --generate_hdf
--dataset imdb-light
--csv_seperator ,
--csv_path ../imdb-benchmark
--hdf_path ../imdb-benchmark/gen_single_light
--max_rows_per_hdf_file 100000000
Generate sampled hdf files from csvs.
python3 maqp.py --generate_sampled_hdfs
--dataset imdb-light
--hdf_path ../imdb-benchmark/gen_single_light
--max_rows_per_hdf_file 100000000
--hdf_sample_size 10000
Learn ensemble with the optimized rdc strategy (requires postgres with imdb dataset)
python3 maqp.py --generate_ensemble
--dataset imdb-light
--samples_per_spn 10000000 10000000 1000000 1000000 1000000
--ensemble_strategy rdc_based
--hdf_path ../imdb-benchmark/gen_single_light
--max_rows_per_hdf_file 100000000
--samples_rdc_ensemble_tests 10000
--ensemble_path ../imdb-benchmark/spn_ensembles
--database_name imdb
--post_sampling_factor 10 10 5 1 1
--ensemble_budget_factor 5
--ensemble_max_no_joins 3
--pairwise_rdc_path ../imdb-benchmark/spn_ensembles/pairwise_rdc.pkl
Alternatively: Learn base ensemble over different tables with naive strategy. (Does not work with different dataset sizes because join sizes are hard coded but does not require postgres)
python3 maqp.py --generate_ensemble
--dataset imdb-light
--samples_per_spn 1000000 1000000 1000000 1000000 1000000
--ensemble_strategy relationship
--hdf_path ../imdb-benchmark/gen_single_light
--ensemble_path ../imdb-benchmark/spn_ensembles
--max_rows_per_hdf_file 100000000
--post_sampling_factor 10 10 5 1 1
Evaluate performance for queries.
python3 maqp.py --evaluate_cardinalities
--rdc_spn_selection
--max_variants 1
--pairwise_rdc_path ../imdb-benchmark/spn_ensembles/pairwise_rdc.pkl
--dataset imdb-light
--target_path ./baselines/cardinality_estimation/results/deepDB/imdb_light_model_based_budget_5.csv
--ensemble_location ../imdb-benchmark/spn_ensembles/ensemble_join_3_budget_5_10000000.pkl
--query_file_location ./benchmarks/job-light/sql/job_light_queries.sql
--ground_truth_file_location ./benchmarks/job-light/sql/job_light_true_cardinalities.csv
Updates
Conditional incremental learning (i.e., initial learning of all films before 2013, newer films learn incremental)
python3 maqp.py --generate_ensemble
--dataset imdb-light
--samples_per_spn 10000000 10000000 1000000 1000000 1000000
--ensemble_strategy rdc_based
--hdf_path ../imdb-benchmark/gen_single_light
--max_rows_per_hdf_file 100000000
--samples_rdc_ensemble_tests 10000
--ensemble_path ../imdb-benchmark/spn_ensembles
--database_name JOB-light
--post_sampling_factor 10 10 5 1 1
--ensemble_budget_factor 0
--ensemble_max_no_joins 3
--pairwise_rdc_path ../imdb-benchmark/spn_ensembles/pairwise_rdc.pkl
--incremental_condition "title.production_year<2013"
Optimized Inference
Generate the C++ code. (Currently only works for cardinality estimation).
python3 maqp.py --code_generation
--ensemble_path ../imdb-benchmark/spn_ensembles/ensemble_join_3_budget_5_10000000.pkl
Compile it in a venv with pybind installed.
Sometimes installing this yields: ModuleNotFoundError: No module named 'pip.req'
One workaround is to downgrade pip pip3 install pip==9.0.3 as described here.
The command below works for ubuntu 18.04. Make sure the generated .so file is in the root directory of the project.
g++ -O3 -Wall -shared -std=c++11 -ftemplate-depth=2048 -ftime-report -fPIC `python3 -m pybind11 --includes` optimized_inference.cpp -o optimized_inference`python3-config --extension-suffix`
If you now want to leverage the module you have to specify it for cardinalities.
python3 maqp.py --evaluate_cardinalities
--rdc_spn_selection
--max_variants 1
--pairwise_rdc_path ../imdb-benchmark/spn_ensembles/pairwise_rdc.pkl
--dataset imdb-light
--target_path ./baselines/cardinality_estimation/results/deepDB/imdb_light_model_based_budget_5.csv
--ensemble_location ../imdb-benchmark/spn_ensembles/ensemble_join_3_budget_5_10000000.pkl
--query_file_location ./benchmarks/job-light/sql/job_light_queries.sql
--ground_truth_file_location ./benchmarks/job-light/sql/job_light_true_cardinalities.csv
--use_generated_code
AQP
SSB pipeline
Generate standard SSB dataset (Scale Factor=500) and use the correct seperator.
for i in `ls *.tbl`; do
sed 's/|$//' $i > $TMP_DIR/${i/tbl/csv} &
echo $i;
done
Create lineorder sample
cat lineorder.csv | awk 'BEGIN {srand()} !/^$/ { if (rand() <= .003333) print $0}' > lineorder_sampled.csv
Generate hdf files from csvs.
python3 maqp.py --generate_hdf
--dataset ssb-500gb
--csv_seperator \|
--csv_path ../mqp-data/ssb-benchmark
--hdf_path ../mqp-data/ssb-benchmark/gen_hdf
Learn the ensemble with a naive strategy.
python3 maqp.py --generate_ensemble
--dataset ssb-500gb
--samples_per_spn 1000000
--ensemble_strategy single
--hdf_path ../mqp-data/ssb-benchmark/gen_hdf
--ensemble_path ../mqp-data/ssb-benchmark/spn_ensembles
--rdc_threshold 0.3
--post_sampling_factor 10
Optional: Compute ground truth for AQP queries (requires postgres with ssb schema).
python3 maqp.py --aqp_ground_truth
--query_file_location ./benchmarks/ssb/sql/aqp_queries.sql
--target_path ./benchmarks/ssb/ground_truth_500GB.pkl
--database_name ssb
Evaluate the AQP queries.
python3 maqp.py --evaluate_aqp_queries
--dataset ssb-500gb
--target_path ./baselines/aqp/results/deepDB/ssb_500gb_model_based.csv
--ensemble_location ../mqp-data/ssb-benchmark/spn_ensembles/ensemble_single_ssb-500gb_1000000.pkl
--query_file_location ./benchmarks/ssb/sql/aqp_queries.sql
--ground_truth_file_location ./benchmarks/ssb/ground_truth_500GB.pkl
Optional: Create the ground truth for confidence interval. (with 10M because we also use 10M samples for the training)
python3 maqp.py --aqp_ground_truth
--query_file_location ./benchmarks/ssb/sql/confidence_queries.sql
--target_path ./benchmarks/ssb/confidence_intervals/confidence_interval_10M.pkl
--database_name ssb
Evaluate the confidence intervals.
python3 maqp.py --evaluate_confidence_intervals
--dataset ssb-500gb
--target_path ./baselines/aqp/results/deepDB/ssb500GB_confidence_intervals.csv
--ensemble_location ../mqp-data/ssb-benchmark/spn_ensembles/ensemble_single_ssb-500gb_1000000.pkl
--query_file_location ./benchmarks/ssb/sql/aqp_queries.sql
--ground_truth_file_location ./benchmarks/ssb/confidence_intervals/confidence_interval_10M.pkl
--confidence_upsampling_factor 300
--confidence_sample_size 10000000
Flights pipeline
Generate flights dataset with scale factor 1 billion using IDEBench and generate a sample using
cat dataset.csv | awk 'BEGIN {srand()} !/^$/ { if (rand() <= .01) print $0}' > dataset_sampled.csv
Generate hdf files from csvs.
python3 maqp.py --generate_hdf
--dataset flights1B
--csv_seperator ,
--csv_path ../mqp-data/flights-benchmark
--hdf_path ../mqp-data/flights-benchmark/gen_hdf
Learn the ensemble.
python3 maqp.py --generate_ensemble
--dataset flights1B
--samples_per_spn 10000000
--ensemble_strategy single
--hdf_path ../mqp-data/flights-benchmark/gen_hdf
--ensemble_path ../mqp-data/flights-benchmark/spn_ensembles
--rdc_threshold 0.3
--post_sampling_factor 10
Optional: Compute ground truth
python3 maqp.py --aqp_ground_truth
--dataset flights1B
--query_file_location ./benchmarks/flights/sql/aqp_queries.sql
--target_path ./benchmarks/flights/ground_truth_1B.pkl
--database_name flights
Evaluate the AQP queries.
python3 maqp.py --evaluate_aqp_queries
--dataset flights1B
--target_path ./baselines/aqp/results/deepDB/flights1B_model_based.csv
--ensemble_location ../mqp-data/flights-benchmark/spn_ensembles/ensemble_single_flights1B_10000000.pkl
--query_file_location ./benchmarks/flights/sql/aqp_queries.sql
--ground_truth_file_location ./benchmarks/flights/ground_truth_1B.pkl
Optional: Create the ground truth for confidence interval. (with 10M because we also use 10M samples for the training)
python3 maqp.py --aqp_ground_truth
--dataset flights1B
--query_file_location ./benchmarks/flights/sql/confidence_queries.sql
--target_path ./benchmarks/flights/confidence_intervals/confidence_interval_10M.pkl
--database_name flights10M_origsample
Evaluate the confidence intervals.
python3 maqp.py --evaluate_confidence_intervals
--dataset flights1B
--target_path ./baselines/aqp/results/deepDB/flights1B_confidence_intervals.csv
--ensemble_location ../mqp-data/flights-benchmark/spn_ensembles/ensemble_single_flights1B_10000000.pkl
--query_file_location ./benchmarks/flights/sql/aqp_queries.sql
--ground_truth_file_location ./benchmarks/flights/confidence_intervals/confidence_interval_10M.pkl
--confidence_upsampling_factor 100
--confidence_sample_size 10000000
TPC-DS (Single Table) pipeline
As an additional example on how to work with DeepDB, we provide an example on just a single table of the TPC-DS schema for the queries in ./benchmarks/tpc_ds_single_table/sql/aqp_queries.sql. As a prerequisite, you need a 10 million tuple sample of the store_sales table in the directory ../mqp-data/tpc-ds-benchmark/store_sales_sampled.csv. Afterwards,
you can run the following commands. To compute the ground truth, you need a postgres instance with a 1T TPC-DS dataset.
Generate hdf files from csvs
python3 maqp.py --generate_hdf
--dataset tpc-ds-1t
--csv_seperator |
--csv_path ../mqp-data/tpc-ds-benchmark
--hdf_path ../mqp-data/tpc-ds-benchmark/gen_hdf
Learn the ensemble
python3 maqp.py --generate_ensemble
--dataset tpc-ds-1t
--samples_per_spn 10000000
--ensemble_strategy single
--hdf_path ../mqp-data/tpc-ds-benchmark/gen_hdf
--ensemble_path ../mqp-data/tpc-ds-benchmark/spn_ensembles
--rdc_threshold 0.3
--post_sampling_factor 10
Compute ground truth
python3 maqp.py --aqp_ground_truth
--dataset tpc-ds-1t
--query_file_location ./benchmarks/tpc_ds_single_table/sql/aqp_queries.sql
--target_path ./benchmarks/tpc_ds_single_table/ground_truth_1t.pkl
--database_name tcpds
Evaluate the AQP queries
python3 maqp.py --evaluate_aqp_queries
--dataset tpc-ds-1t
--target_path ./baselines/aqp/results/deepDB/tpcds1t_model_based.csv
--ensemble_location ../mqp-data/tpc-ds-benchmark/spn_ensembles/ensemble_single_tpc-ds-1t_10000000.pkl
--query_file_location ./benchmarks/tpc_ds_single_table/sql/aqp_queries.sql
--ground_truth_file_location ./benchmarks/tpc_ds_single_table/ground_truth_1t.pkl
How to experiment with DeepDB on a new Dataset
- Specify a new schema in the schemas folder
- Due to the current implementation, make sure to declare
- the primary key,
- the filename of the csv sample file,
- the correct table size and sample rate,
- the relationships among tables if you do not just run queries over a single table,
- any non-key functional dependencies (this is rather an implementation detail),
- and include all columns in the no-compression list by default (as done for the IMDB benchmark),
- To further reduce the training time, you can exclude columns you do not need in your experiments (also done in the IMDB benchmark)
- Generate the HDF/sampled HDF files and learn the RSPN ensemble
- Use the RSPN ensemble to answer queries
- For reference, please check the commands to reproduce the results of the paper
Reference
If you find this repository useful in your work, please cite our paper:
@inproceedings{deepdb,
title={DeepDB: Learn from Data, not from Queries!},
author = {Benjamin Hilprecht and Andreas Schmidt and Moritz Kulessa and Alejandro Molina and Kristian Kersting and Carsten Binnig},
journal={Proceedings of the VLDB Endowment},
volume={13},
number={7},
pages={992--1005},
year={2020},
publisher={VLDB Endowment}
}