CardinalityEstimationTestbed

The source code of Paper Learned Cardinality Estimation: A Design Space Exploration and A Comparative Evaluation.

Experiment of synthetic

Generate_data_sql

cd CardinalityEstimationTestbed/Synthetic
python generate_data_sql.py --cols [COLUMNS_NUM] --distinct [DOMAIN_SIZE] --corr [CORRELATION] --skew [SKEWNESS]

Get_sql_truecard

python get_truecard.py --version cols_[COLUMNS_NUM]_distinct_[DOMAIN_SIZE]_corr_[CORRELATION]_skew_[SKEWNESS]

Get_result

python get_result.py --cols [COLUMNS_NUM] --distinct [DOMAIN_SIZE] --corr [CORRELATION] --skew [SKEWNESS] --method [METHOD]

• Then experimental results of all methods can be obtained.

Parameters

• [COLUMNS_NUM] should be an integer between 1 and 9.
• [DOMAIN_SIZE] should be an integer between 10 and 20000.
• [CORRELATION] should be an integer between 1 and 9.
• [SKEWNESS] should be an integer between 1 and 9.
• The parameters we use is

cols in [2, 4, 6, 8]; distinct in [10, 100, 1000, 10000]; corr in [2, 4, 6, 8]; skew in [2, 4, 6, 8];

Experiment of overall

Real Datasets prepare

• Download the IMDB data tables to: ../train-test-data/imdbdataset-str
• Download the forest, power data tables to: ../train-test-data/forest_power-data-sql
• We open source XTZX. We have put the data and workload for the experiment into the repositorie. Begin by doing a simple job on the table, removing some unused columns. For Forest We use first 10 numeric attributes; For Power We used the 7 numeric attributes after the first two attributes (date and time).
• Remember to put the tables in their respective folders that do not contain table headers, for example: forest_power-data-sql/no_head, imdbdata-num/no_head, xtzx-data-sql/no_head.

Forest & Power

• Run the forest.sh, power.sh in each method folder to execute the code to get the result.

Imdb

• First, the strings in the data tables should be converted to numbers.
  cd CardinalityEstimationTestbed/Overall/imdb
python data_str2num.py

Varying Columns

• cd Overall/[METHOD]
python run_exp.py --cols [cols]
• For example:
  cd Overall/deepdb
python run_exp.py --cols 4
  AND cd Overall/xgboost_localnn
python run_exp.py --cols 4 --model [nn OR xgb]

Varying Domain Size

• First, the data of the distinct is generated and populated.
  cd CardinalityEstimationTestbed/Overall/imdb/distinct
python data_process.py
• Then refer to run.sh in mscn and neurocard folder to execute the code to get the result.

Varying Join samples

• For example:
  cd Overall/deepdb
python run_exp.py --cols 4 --samples 3

XTZX

• cd Overall/[METHOD]
python run_exp.py --xt [cols]
• For example:
  cd Overall/deepdb
python run_exp.py --xt 4
  AND cd Overall/xgboost_localnn
python run_exp.py --xt 4 --model [nn OR xgb]

Update

• Select 10% of the number of rows in each table and add them to the end of the table.
• Refer to update.sh in each method folder to execute the code to get the result.

Citation

@article{DBLP:journals/pvldb/benchmark,
	author    = {Ji Sun and Jintao Zhang and Zhaoyan Sun and Guoliang Li and Nan Tang},
	title     = {Learned Cardinality Estimation: A Design Space Exploration and A Comparative Evaluation},
	journal   = {VLDB},
	year      = {2021},
}