XMUDM / TRAP

TRAP

This is the code respository of TRAP, which is a unified framework concerning the robustness assessment over various index advisors. Specifically, the assessment is conducted based on the perturbation-based adversarial workloads according to the observations from the open-source benchmarks and real world workloads.

The repository contains the following contents:

1. Running Script: The scripts for the experiment setup and the workload generation.

2. Constraint-Aware Reference Tree: The additional information, i.e., the BNF Grammar, the token type and the vocabulary.

3. Cost estimation: The cost estimation results, i.e., the learned model's accuaracy over different models with various feature channels.

1.1 Setup

We introduce the indispensable step, i.e., experiment setup before workload generation in TRAP, you should check the following things:

• Create the TPC-H/TPC-DS database instance according to the toolkit provided in TPC-Homepage;
• Create the HypoPG extension on the TPC-H/TPC-DS database instance for the usage of hypothetical index according to HypoPG/hypopg: Hypothetical Indexes for PostgreSQL (github.com);
• Create the python virtual environment tailored for TRAP. Specifically, you can utilize the following script and the corresponding file requirements.txt is provided under the main directory. Please check the packages required are properly installed.

# Create the virtualenv `TRAP`
conda create -n TRAP python=3.7		 	

# Activate the virtualenv `TRAP`
conda activate TRAP				

# Install requirements with pip
while read requirement; do pip install $requirement; done < requirements.txt	

1.2 Module List

The description of files in each module is shown below.

Module	Description
data_resource	The configuration over database (e.g., the user name), heuristic-based and learning-based index advisors (e.g., the tuning constraint), the global SQL token's vocabulary and the sample data leveraged during the generation of workload.
heuristic_advisor	The implementation of heuristic-based index advisors.
learning_advisor	The implementation of learning-based index advisors.
workload_generation	The implementation of adversarial workload generation module.

1.3 Adversarial Workload Generation

You should make sure the packages required in the requirements.txt are properly installed and connect to your own database instance to specify the configuration in /data_source/database_conf/db_info.conf!!!

python main.py 
--exp_id gen_exp_id --gpu_no -1 --seed 666  
--is_bid --is_attn --rnn_type GRU --model_struct Seq2Seq 

--rein_epoch 10 --rein_lr 0.001 --train_mode rl_pg 
--max_diff 5 --pert_mode all 
--reward dynamic --reward_form cost_red_ratio 

--victim extend --sel_param parameters 
--exp_file ./data_resource/heuristic_conf/extend_config.json  

--db_file ./data_resource/database_conf/db_info.conf 
--data_load ./data_resource/sample_data/sample_data.pt 

--model_load ./data_resource/Pre-train_100.pt  
--colinfo_file ./data_resource/database_conf/colinfo.json 
--wordinfo_file ./data_resource/vocab/wordinfo.json 
--schema_file ./data_resource/database_conf/schema.json  

--batch_size 32 --dropout 0.5 
--enc_hidden_size 128 --dec_hidden_size 128 
--src_vbs 3040 --tgt_vbs 3040

2. BNF Grammar, Token Type and Vocabulary

The BNF Grammar for a standard Select-Project-Aggregate-Join (SPAJ) query is presented in the following table. Please refer to the directory, i.e., /data_source/vocab/* for the details of the token type and the vocabulary.

SQL ::= SELECT FROM [WHERE] [GROUPBY] [HAVING] [ORDERBY] SELECT ::= "select" (term ("," term)? | SQL) FROM ::= "from" (table (("," table)? | ("join" table)?) | SQL) WHERE ::= "where" predicate1 (conj predicate1)? GROUPBY ::= "group by" column ("," column)? HAVING ::= "having" predicate2 ("," predicate2)? ORDERBY ::= "order by" (term | term key) ("," term | "," term key)?
term ::= column | agg "(" column ")" predicate1 ::= column op (value | SQL) predicate2 ::= term op (value | SQL) table ::= "t1" | "t2" column ::= "t1.col1" | "t1.col2" | "t2.col1" agg ::= "min" | "max" | "count" | "sum" | "avg" op ::= ">" | "<" | ">=" | "<=" | "!=" conj ::= "and" | "or"value ::= "value1" | "value2" key ::= "asc" | "desc"