mongo/buildscripts/cost_model/README.md
Steve McClure 32e8f260de SERVER-124136 Format markdown via prettier: wrap lines and use width of 100 (#52231)
GitOrigin-RevId: 3305c1e2ee3a6a2c3a5b2b7883b0f491a59ed646
2026-04-21 19:20:11 +00:00

2.3 KiB

Cost Model Calibrator

Getting Started

1) Setup Mongod

First, prepare the MongoDB server:

  1. Activate the standard virtual environment:
source python3-venv/bin/activate
  1. Build server with optimizations (makes doc insertion faster):
(python3-venv) bazel build --config=opt install-devcore
  1. Run mongod instance (only for CBR calibration, because join_start.py manages mongod's lifecycle itself):
(python3-venv) bazel-bin/install-mongod/bin/mongod --setParameter internalMeasureQueryExecutionTimeInNanoseconds=true

2) Setup Cost Model Calibrator

In another terminal:

  1. Navigate to the cost model directory:
cd buildscripts/cost_model
  1. Set up Python alias to use MongoDB toolchain:
alias python=/opt/mongodbtoolchain/v4/bin/python3
  1. Deactivate any existing Python environment (if needed):
deactivate
  1. Create new virtual environment:
/opt/mongodbtoolchain/v4/bin/python3 -m venv cm
  1. Activate the new environment:
source cm/bin/activate
  1. Install required packages:
(cm) python -m pip install -r requirements.txt
  1. Run the calibrator:
  • For CBR cost model calibration:
    (cm) python start.py
    
  • For JOO cost model calibration:
    (cm) python join_start.py
    
    To skip the constant calibration (warm scan, CPU, sequential I/O, random I/O) and only run the join algorithm comparison:
    (cm) python join_start.py --join-only
    
    To iterate quickly on cost model changes, reuse pre-recorded execution times from a previous full run. This skips actual query execution, only running queryPlanner explains to collect fresh cost estimates:
    (cm) python join_start.py --execution-times join_output/join_times_in-cache.csv join_output/join_times_exceeds-cache.csv
    

Note: For CBR calibration, the first time it will take a while since it has to generate the data. Afterwards, as long as you aren't modifying the collections, you can comment out await generator.populate_collections() in start.py - this will make it a lot faster.

  1. When done, deactivate the environment:
(cm) deactivate

Install New Packages

  1. Install the package:
(cm) python -m pip install <package_name>
  1. Update requirements.txt:
(cm) python -m pip freeze > requirements.txt