mongo/buildscripts/cost_model/qsn_calibrator.py

# Copyright (C) 2022-present MongoDB, Inc.
#
# This program is free software: you can redistribute it and/or modify
# it under the terms of the Server Side Public License, version 1,
# as published by MongoDB, Inc.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
# Server Side Public License for more details.
#
# You should have received a copy of the Server Side Public License
# along with this program. If not, see
# <http://www.mongodb.com/licensing/server-side-public-license>.
#
# As a special exception, the copyright holders give permission to link the
# code of portions of this program with the OpenSSL library under certain
# conditions as described in each individual source file and distribute
# linked combinations including the program with the OpenSSL library. You
# must comply with the Server Side Public License in all respects for
# all of the code used other than as permitted herein. If you modify file(s)
# with this exception, you may extend this exception to your version of the
# file(s), but you are not obligated to do so. If you do not wish to do so,
# delete this exception statement from your version. If you delete this
# exception statement from all source files in the program, then also delete
# it in the license file.
#
"""Calibrate QSN nodes."""

from __future__ import annotations

import os

import experiment as exp
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import statsmodels.api as sm
from config import QsNodeCalibrationConfig, QuerySolutionCalibrationConfig
from cost_estimator import estimate
from database_instance import DatabaseInstance
from sklearn.linear_model import LinearRegression

__all__ = ["calibrate"]


async def calibrate(config: QuerySolutionCalibrationConfig, database: DatabaseInstance):
    """Main entry-point for QSN calibration."""

    if not config.enabled:
        return {}

    df = await exp.load_calibration_data(database, config.input_collection_name)
    noout_df = exp.remove_outliers(df, 0.1, 0.90)
    qsn_df = exp.extract_qsn_nodes(noout_df)
    result = {}
    for node_config in config.nodes:
        key = node_config.name if node_config.name else node_config.type
        result[key] = calibrate_node(qsn_df, config, node_config)
    return result


def calibrate_node(
    qsn_df: pd.DataFrame,
    config: QuerySolutionCalibrationConfig,
    node_config: QsNodeCalibrationConfig,
):
    node_name = node_config.name if node_config.name else node_config.type
    qsn_node_df = qsn_df[
        (qsn_df.node_type == node_config.type) & (qsn_df.note.isna() | (qsn_df.note == node_name))
    ]
    if node_config.filter_function is not None:
        qsn_node_df = node_config.filter_function(qsn_node_df)
    y = qsn_node_df["execution_time"]
    X_vars = (
        pd.DataFrame({"Number of documents": qsn_node_df["n_processed"]})
        if node_config.variables_override is None
        else node_config.variables_override(qsn_node_df)
    )
    n_vars = X_vars.shape[1]
    labels = X_vars.columns.tolist()
    X = sm.add_constant(X_vars)

    # TODO(SERVER-110398): Clean this up by not relying on fixed '2' value
    if node_config.type == "SORT" and not X_vars.empty and "SPILL" not in node_name:
        # This prints the average execution time for input cardinality '2'
        mask = X_vars.iloc[:, 0] == 2.0
        avg_exec_time = y[mask].mean()
        print(f"{node_name}: Startup Cost -> {avg_exec_time} ns")

    def fit(X, y):
        nnls = LinearRegression(positive=True, fit_intercept=False)
        model = nnls.fit(X, y)
        return (model.coef_, model.predict)

    model = estimate(fit, X.to_numpy(), y.to_numpy(), config.test_size, config.trace)
    # plot regression and save to file
    if model.predict:
        if n_vars == 2:
            fig = plt.figure()
            ax = fig.add_subplot(111, projection="3d")
            X1 = X_vars.iloc[:, 0]
            X2 = X_vars.iloc[:, 1]
            ax.scatter(X1, X2, y, label="Executions", color="blue")

            x1_range = np.linspace(X1.min(), X1.max())
            x2_range = np.linspace(X2.min(), X2.max())
            x1_mesh, x2_mesh = np.meshgrid(x1_range, x2_range)

            ax.plot_surface(
                x1_mesh,
                x2_mesh,
                model.coef[0] * x1_mesh + model.coef[1] * x2_mesh + model.intercept,
                alpha=0.5,
                color="orange",
                label="Regression Plane",
            )

            ax.set(
                xlabel=labels[0],
                ylabel=labels[1],
                zlabel="Execution time (ns)",
                title=f"Regression for {node_name}",
            )
            ax.legend()
        elif n_vars == 1:
            fig, ax = plt.subplots()
            ax.scatter(X_vars, y, label="Executions")
            ax.plot(
                X_vars,
                model.predict(X),
                linewidth=3,
                color="tab:orange",
                label="Linear Regression",
            )
            ax.set(
                xlabel=labels[0],
                ylabel="Execution time (ns)",
                title=f"Regression for {node_name}",
            )

            ax.legend()
        else:
            # For 3 variables or more, we just plot the predicted vs actual execution time
            y_pred = model.predict(X)
            fig, ax = plt.subplots()
            ax.scatter(y_pred, y, alpha=0.6)
            min_val = min(y.min(), y_pred.min())
            max_val = max(y.max(), y_pred.max())
            ax.plot([min_val, max_val], [min_val, max_val], "r--", lw=2)
            ax.set_xlabel("Predicted Execution Time (ns)")
            ax.set_ylabel("Actual Execution Time (ns)")
            ax.set_title(f"Predicted vs Actual for {node_name}")
            plt.tight_layout()

        if fig:
            os.makedirs("figures", exist_ok=True)
            fig.savefig("figures/" + f"{node_name}.png")
    return model