feat: add ab_plot.py script

Add a script to create visualizations of performance data.
- works for both single and multiple inputs
- allows comparison of different hosts
- allows comparison of multiple tests
- can generate pdf and html outputs

Signed-off-by: Egor Lazarchuk <[email protected]>

Author: ShadowCurse

tests/README.md

@@ -251,6 +251,15 @@ schedule an A/B-Test in buildkite, the `REVISION_A` and `REVISION_B` environment
 variables need to be set in the "Environment Variables" field under "Options" in
 buildkite's "New Build" modal.
 
+### A/B visualization
+
+To create visualization of A/B runs use `tools/ab_plot.py` script. Example
+usage:
+
+```sh
+ ./tools/plot.py a_path b_path --output_type pdf
+```
+
 ### Beyond commit comparisons
 
 While our automated A/B-Testing suite only supports A/B-Tests across commit

tools/ab_plot.py

@@ -0,0 +1,310 @@
+#!/usr/bin/env python3
+# Copyright 2025 Amazon.com, Inc. or its affiliates. All Rights Reserved.
+# SPDX-License-Identifier: Apache-2.0
+"""
+Script for creating visualizations for A/B runs.
+
+Usage:
+ab_plot.py path_to_run_a path_to_run_b path_to_run_c ... --output_type pdf/table
+"""
+
+import argparse
+import glob
+import json
+import time
+from pathlib import Path
+from typing import Callable, List
+
+import matplotlib.pyplot as plt
+import numpy as np
+import pandas as pd
+import scipy
+import seaborn as sns
+from matplotlib.backends.backend_pdf import PdfPages
+
+pd.set_option("display.float_format", "{:.2f}".format)
+
+
+def check_regression(
+    a_samples: List[float],
+    b_samples: List[float],
+    statistic: Callable = np.mean,
+    *,
+    n_resamples=9999,
+):
+    """
+    Check if 2 sample groups have a statistically big enough difference
+    """
+    result = scipy.stats.permutation_test(
+        (a_samples, b_samples),
+        lambda x, y: statistic(y) - statistic(x),
+        vectorized=False,
+        n_resamples=n_resamples,
+    )
+    statistic_a = statistic(a_samples)
+
+    return result.pvalue, result.statistic / statistic_a, result.statistic
+
+
+def load_data(data_path: Path):
+    """
+    Recursively collects `metrics.json` files in provided path
+    """
+    data = []
+    for name in glob.glob(f"{data_path}/**/metrics.json", recursive=True):
+        with open(name, encoding="utf-8") as f:
+            j = json.load(f)
+
+        if "performance_test" not in j["dimensions"]:
+            print(f"skipping: {name}")
+            continue
+
+        metrics = j["metrics"]
+        # Move test name from dimentions into a separate column
+        perf_test = j["dimensions"]["performance_test"]
+        del j["dimensions"]["performance_test"]
+
+        # These are host specific and will prevent comparison of
+        # different hosts
+        del j["dimensions"]["instance"]
+        del j["dimensions"]["cpu_model"]
+        del j["dimensions"]["host_kernel"]
+
+        dimentions = frozenset(j["dimensions"].items())
+
+        for m in metrics:
+            if "cpu_utilization" in m:
+                continue
+            mm = metrics[m]
+            unit = mm["unit"]
+            values = mm["values"]
+            for i, v in enumerate(values):
+                data.append(
+                    {
+                        "index": i,
+                        "test": perf_test,
+                        "metric": m,
+                        "value": v,
+                        "unit": unit,
+                        "dimensions": dimentions,
+                    }
+                )
+
+    return data
+
+
+def p50(a):
+    """Returns 50th percentile of 1d-array a"""
+    return np.percentile(a, 50)
+
+
+def p90(a):
+    """Returns 90th percentile of 1d-array a"""
+    return np.percentile(a, 90)
+
+
+def create_table(df: pd.DataFrame):
+    """Create an html table per test in the data frame"""
+
+    for test_value in df["test"].unique():
+        df_test = df[df["test"] == test_value]
+
+        # Split dimensions into separate columns
+        df_expanded = df_test.copy()
+        dim_data = []
+        for _, row in df_expanded.iterrows():
+            dim_dict = dict(row["dimensions"])
+            dim_data.append(dim_dict)
+
+        # Need to reset indexes because otherwise `pd.concat` will add NaN in all
+        # rows where indexes differ
+        dim_df = pd.DataFrame(dim_data).reset_index(drop=True)
+        df_data = df_expanded.drop("dimensions", axis=1).reset_index(drop=True)
+        df_expanded = pd.concat([df_data, dim_df], axis=1)
+
+        # Use dimension columns as index
+        dim_cols = sorted(list(dim_df.columns))
+        df_pivoted = df_expanded.pivot_table(
+            values=["value"],
+            index=["metric", "unit"] + dim_cols,
+            columns="group",
+            aggfunc=[p50, p90],
+        )
+
+        # Add comparison columns for each group vs first group (A)
+        groups = sorted(df_test["group"].unique())
+        for baseline in groups:
+            for group in groups:
+                if group == baseline:
+                    continue
+                for stat in ["p50", "p90"]:
+                    diff_col = (stat, "value", f"{baseline}->{group} %")
+                    df_pivoted[diff_col] = (
+                        (
+                            df_pivoted[(stat, "value", group)]
+                            - df_pivoted[(stat, "value", baseline)]
+                        )
+                        / df_pivoted[(stat, "value", baseline)]
+                        * 100.0
+                    )
+                    diff_col = (stat, "value", f"{baseline}->{group} abs")
+                    df_pivoted[diff_col] = (
+                        df_pivoted[(stat, "value", group)]
+                        - df_pivoted[(stat, "value", baseline)]
+                    )
+
+        # Sort columns to have a persistent table representation
+        df_pivoted = df_pivoted[sorted(df_pivoted.columns)]
+
+        test_output_path = f"{test_value}.html"
+        with open(test_output_path, "w", encoding="UTF-8") as writer:
+            writer.write("<br>")
+            styled = df_pivoted.style.format(precision=2)
+            styled = styled.set_table_attributes("border=1")
+            styled = styled.set_table_styles(
+                [{"selector": 'th:contains("->")', "props": [("min-width", "80px")]}]
+            )
+
+            # Apply color gradient to all comparison columns
+            for baseline in groups:
+                for group in groups:
+                    if group == baseline:
+                        continue
+                    for stat in ["p50", "p90"]:
+                        diff_col = (stat, "value", f"{baseline}->{group} %")
+                        styled = styled.background_gradient(
+                            subset=[diff_col], cmap="RdYlGn"
+                        )
+
+            writer.write(styled.to_html())
+            writer.write("<br>")
+        print(f"Ready: {test_output_path}")
+
+
+def create_pdf(args, df: pd.DataFrame):
+    """Create a pdf per test in the data frame"""
+
+    sns.set_style("whitegrid")
+    metrics = df["metric"].unique()
+    n_groups = len(df["group"].unique())
+
+    for test_value in df["test"].unique():
+        test_output_path = f"{test_value}.pdf"
+        with PdfPages(test_output_path) as pdf:
+            df_test = df[df["test"] == test_value]
+            for dim_value in df_test["dimensions"].unique():
+                for metric in metrics:
+                    metric_data = df_test[
+                        (df_test["metric"] == metric)
+                        & (df_test["dimensions"] == dim_value)
+                    ]
+
+                    if len(metric_data) == 0:
+                        continue
+
+                    additional_title = ""
+                    fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(15, 6))
+                    if n_groups == 2:
+                        # Check if difference is significant
+                        a_values = metric_data[metric_data["group"] == "A"][
+                            "value"
+                        ].values
+                        b_values = metric_data[metric_data["group"] == "B"][
+                            "value"
+                        ].values
+                        pvalue, diff_rel, diff_abs = check_regression(
+                            a_values, b_values
+                        )
+
+                        if (
+                            pvalue <= 0.1
+                            and abs(diff_rel) >= 0.05
+                            and abs(diff_abs) >= 0.0
+                        ):
+                            fig.patch.set_facecolor("lightcoral")
+                            additional_title = (
+                                f"{diff_rel * 100:+.2f}% ({diff_abs:+.2f}) difference"
+                            )
+
+                    # Make a multi-line title since single line will be too long
+                    dim_items = sorted(str(item) for item in dim_value)
+                    dim_chunks = [
+                        ", ".join(dim_items[i : i + 4])
+                        for i in range(0, len(dim_items), 4)
+                    ]
+                    dim_str = "\n".join(dim_chunks)
+                    title = f"{metric}\n{dim_str}\n{additional_title}"
+                    if additional_title:
+                        weight = "bold"
+                    else:
+                        weight = "normal"
+                    fig.suptitle(title, fontsize=10, weight=weight)
+
+                    sns.boxenplot(data=metric_data, x="group", y="value", ax=ax1)
+                    ax1.set_ylabel(f"{metric} ({metric_data['unit'].iloc[0]})")
+
+                    metric_data_indexed = metric_data.reset_index()
+                    errorbar = (args.errorbar[0], int(args.errorbar[1]))
+                    sns.lineplot(
+                        data=metric_data_indexed,
+                        x="index",
+                        y="value",
+                        hue="group",
+                        ax=ax2,
+                        errorbar=errorbar,
+                    )
+                    ax2.set_ylabel(f"{metric} ({metric_data['unit'].iloc[0]})")
+
+                    plt.tight_layout()
+                    pdf.savefig()
+                    plt.close()
+        print(f"Ready: {test_output_path}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="Executes Firecracker's A/B testsuite across the specified commits"
+    )
+    parser.add_argument(
+        "paths",
+        nargs="+",
+        help="Paths to directories with test runs",
+        type=Path,
+    )
+    parser.add_argument(
+        "--errorbar",
+        nargs=2,
+        default=["pi", "95"],
+        help="Errorbar configuration for lineplot (type, value)",
+    )
+    parser.add_argument(
+        "--output_type",
+        default=["pdf"],
+        help="Type of the output to generate",
+    )
+    args = parser.parse_args()
+
+    # Data retrieval
+    start_time = time.time()
+    all_data = []
+    for i, path in enumerate(args.paths):
+        data = load_data(path)
+        print(f"getting data {i} from {path}: {len(data)}")
+        df = pd.DataFrame(data)
+        df["group"] = chr(65 + i)  # A, B, C, D, ...
+        all_data.append(df)
+    print(f"Data retrieval: {time.time() - start_time:.2f}s")
+
+    # Data processing
+    start_time = time.time()
+    df_combined = pd.concat(all_data, ignore_index=True)
+    print(f"Data processing: {time.time() - start_time:.2f}s")
+
+    # Plotting
+    start_time = time.time()
+    if args.output_type == "pdf":
+        create_pdf(args, df_combined)
+    if args.output_type == "table":
+        create_table(df_combined)
+
+    print(f"Plotting: {time.time() - start_time:.2f}s")