Skip to content

README.md

Latest commit

 

History

History
137 lines (110 loc) · 4.8 KB

README.md

File metadata and controls

137 lines (110 loc) · 4.8 KB

λ-Tune

The source code of λ-Tune: A Database System Tuning framework based on Large Language Models.

λ-Tune was presented at ACM SIGMOD 2025, Berlin, Germany.

Paper Link: https://dl.acm.org/doi/10.1145/3709652

Prerequisites

Ensure you have Python installed on your system. The script is written in Python and requires necessary permissions to execute.

Database System

Provide the credentials of the target database system (Postgres or MySQL) in lambdatune/resources/config.ini:

[LAMBDA_TUNE]
llm = gpt-4
openai_key = <your-openai-key>
anthropic_key =          ; optional — for Anthropic models

[POSTGRES]
user = <your-pg-user>

Note: config.ini is gitignored and will never be committed.

Install Dependencies

MacOS

brew install pkg-config mysql-client
export PKG_CONFIG_PATH="/opt/homebrew/opt/mysql-client/lib/pkgconfig"

virtualenv .venv
source .venv/bin/activate
pip install -r requirements.txt

Usage

PYTHONPATH=$PWD python lambdatune/run_lambdatune.py --configs $CONFIGS_DIR --out $OUTPUT_FOLDER --system $DBMS

Where $CONFIGS_DIR is the folder with LLM-generated configurations, $OUTPUT_FOLDER is where benchmark results are saved, and $DBMS is the database system to tune (POSTGRES or MYSQL).

Arguments

--benchmark BENCHMARK     Benchmark to run: tpch (default), tpcds, job
--system SYSTEM           Database system: POSTGRES (default), MYSQL
--configs CONFIGS         Path to configs dir, or new dir name when using --config_gen
--out OUT                 Results output directory
--config_gen              Generate new configurations via LLM
--num-configs N           Number of configurations to generate (default: 3)
--indexes-only            Generate index recommendations only, skip system knobs
--cores CORES             Number of CPU cores (default: auto-detected)
--memory MEMORY           Amount of memory in GB (default: auto-detected)

--provider PROVIDER       LLM provider: openai (default), anthropic, ollama, bedrock
--model MODEL             Model name or full LiteLLM string (see examples below)
--api-key API_KEY         API key for the provider (overrides config.ini and env vars)

LLM Providers

λ-Tune uses LiteLLM and supports any provider it covers. The --provider and --model flags override config.ini at runtime.

Provider --provider --model Auth
OpenAI openai gpt-4o, gpt-4 OPENAI_API_KEY or openai_key in config.ini
Anthropic anthropic claude-sonnet-4-6 ANTHROPIC_API_KEY or anthropic_key in config.ini
Ollama (local) ollama llama3, mistral None
AWS Bedrock bedrock anthropic.claude-sonnet-4-5 AWS env vars

You can also pass a fully qualified LiteLLM model string directly via --model without --provider:

--model anthropic/claude-sonnet-4-6

Examples

  1. Run TPC-H over Postgres using an existing configuration directory:

    PYTHONPATH=$PWD python lambdatune/run_lambdatune.py \
  --configs ./lambdatune/configs/tpch_postgres_1 \
  --out ./test \
  --system POSTGRES

  2. Generate new configurations via OpenAI and run the Join Order Benchmark:

    PYTHONPATH=$PWD python lambdatune/run_lambdatune.py \
  --configs new_config --out ./test \
  --system POSTGRES --benchmark job \
  --config_gen --num-configs 5

  3. Use Anthropic Claude instead of OpenAI:

    PYTHONPATH=$PWD python lambdatune/run_lambdatune.py \
  --configs new_config --out ./test \
  --system POSTGRES --config_gen \
  --provider anthropic --model claude-sonnet-4-6 \
  --api-key $ANTHROPIC_API_KEY

  4. Use a local Ollama model (no API key required):

    PYTHONPATH=$PWD python lambdatune/run_lambdatune.py \
  --configs new_config --out ./test \
  --system POSTGRES --config_gen \
  --provider ollama --model llama3

  5. Generate index recommendations only (no system knobs):

    PYTHONPATH=$PWD python lambdatune/run_lambdatune.py \
  --configs new_config --out ./test \
  --system POSTGRES --config_gen --indexes-only

Citation

@article{giannakouris2025lambda,
  title={$\lambda$-tune: Harnessing large language models for automated database system tuning},
  author={Giannakouris, Victor and Trummer, Immanuel},
  journal={Proceedings of the ACM on Management of Data},
  volume={3},
  number={1},
  pages={1--26},
  year={2025},
  publisher={ACM New York, NY, USA}
}