ab_test.py

#!/usr/bin/env python3
# Copyright 2023 Amazon.com, Inc. or its affiliates. All Rights Reserved.
# SPDX-License-Identifier: Apache-2.0
"""
Script for running A/B-Tests

The script takes two git revisions and a pytest integration test. It utilizes
our integration test frameworks --binary-dir parameter to execute the given
test using binaries compiled from each revision, and runs a regression test
comparing resulting metrics between runs.

It performs the A/B-test as follows:
For both A and B runs, collect all `metrics.json` files and read all dimentions
from them. Script assumes all dimentions are unique within single run and both
A and B runs result in the same dimentions. After collection is done, perform
statistical regression test across all the list-valued properties collected.
"""

import argparse
import glob
import json
import os
import statistics
import subprocess
from collections import defaultdict
from pathlib import Path
from typing import Callable, List, Optional, TypeVar

import scipy

UNIT_REDUCTIONS = {
    "Microseconds": "Milliseconds",
    "Milliseconds": "Seconds",
    "Bytes": "Kilobytes",
    "Kilobytes": "Megabytes",
    "Megabytes": "Gigabytes",
    "Gigabytes": "Terabytes",
    "Bits": "Kilobits",
    "Kilobits": "Megabits",
    "Megabits": "Gigabits",
    "Gigabits": "Terabit",
    "Bytes/Second": "Kilobytes/Second",
    "Kilobytes/Second": "Megabytes/Second",
    "Megabytes/Second": "Gigabytes/Second",
    "Gigabytes/Second": "Terabytes/Second",
    "Bits/Second": "Kilobits/Second",
    "Kilobits/Second": "Megabits/Second",
    "Megabits/Second": "Gigabits/Second",
    "Gigabits/Second": "Terabits/Second",
}
INV_UNIT_REDUCTIONS = {v: k for k, v in UNIT_REDUCTIONS.items()}


UNIT_SHORTHANDS = {
    "Seconds": "s",
    "Microseconds": "μs",
    "Milliseconds": "ms",
    "Bytes": "B",
    "Kilobytes": "KB",
    "Megabytes": "MB",
    "Gigabytes": "GB",
    "Terabytes": "TB",
    "Bits": "Bit",
    "Kilobits": "KBit",
    "Megabits": "MBit",
    "Gigabits": "GBit",
    "Terabits": "TBit",
    "Percent": "%",
    "Count": "",
    "Bytes/Second": "B/s",
    "Kilobytes/Second": "KB/s",
    "Megabytes/Second": "MB/s",
    "Gigabytes/Second": "GB/s",
    "Terabytes/Second": "TB/s",
    "Bits/Second": "Bit/s",
    "Kilobits/Second": "KBit/s",
    "Megabits/Second": "MBit/s",
    "Gigabits/Second": "GBit/s",
    "Terabits/Second": "TBit/s",
    "Count/Second": "Hz",
    "None": "",
}


def reduce_value(value, unit):
    """
    Utility function for expressing a value in the largest possible unit in which it would still be >= 1

    For example, `reduce_value(1_000_000, Bytes)` would return (1, Megabytes)
    """
    # Could do this recursively, but I am worried about infinite recursion
    # due to precision problems (e.g. infinite loop of dividing/multiplying by 1000, alternating
    # between values < 1 and >= 1000).
    while abs(value) < 1 and unit in INV_UNIT_REDUCTIONS:
        value *= 1000
        unit = INV_UNIT_REDUCTIONS[unit]
    while abs(value) >= 1000 and unit in UNIT_REDUCTIONS:
        value /= 1000
        unit = UNIT_REDUCTIONS[unit]

    return value, unit


def format_with_reduced_unit(value, unit):
    """
    Utility function for pretty printing a given value by choosing a unit as large as possible,
    and then outputting its shorthand.

    For example, `format_with_reduced_unit(1_000_000, Bytes)` would return "1MB".
    """
    reduced_value, reduced_unit = reduce_value(value, unit)
    formatted_unit = UNIT_SHORTHANDS.get(reduced_unit, reduced_unit)

    return f"{reduced_value:.2f}{formatted_unit}"


# Performance tests that are known to be unstable and exhibit variances of up to 60% of the mean
IGNORED = [
    # Network throughput on m6a.metal
    {"instance": "m6a.metal", "performance_test": "test_network_tcp_throughput"},
    # Network throughput on m7a.metal
    {"instance": "m7a.metal-48xl", "performance_test": "test_network_tcp_throughput"},
    # block latencies if guest uses async request submission
    {"fio_engine": "libaio", "metric": "clat_read"},
    {"fio_engine": "libaio", "metric": "clat_write"},
    # boot time metrics
    {"performance_test": "test_boottime", "metric": "resume_time"},
    # block throughput on m8g
    {"fio_engine": "libaio", "vcpus": "2", "instance": "m8g.metal-24xl"},
    {"fio_engine": "libaio", "vcpus": "2", "instance": "m8g.metal-48xl"},
    # memory hotplug metrics: ignore api_time and fc_time metrics, keeping only total_time.
    *[
        {
            "performance_test": "test_memory_hotplug_latency",
            "metric": f"{prefix}_{metric}",
        }
        for prefix in ["hotplug", "hotunplug", "hotplug_2nd"]
        for metric in ["api_time", "fc_time"]
    ],
]


def is_ignored(dimensions) -> bool:
    """Checks whether the given dimensions match an entry in the IGNORED dictionary above"""
    for high_variance in IGNORED:
        matching = {key: dimensions[key] for key in high_variance if key in dimensions}

        if matching == high_variance:
            return True

    return False


def load_data_series(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)

        metrics = j["metrics"]
        dimentions = frozenset(j["dimensions"].items())

        data[dimentions] = {}
        for m in metrics:
            # Ignore certain metrics as we know them to be volatile
            if "cpu_utilization" in m:
                continue
            mm = metrics[m]
            unit = mm["unit"]
            values = mm["values"]
            data[dimentions][m] = (values, unit)

    return data


def uninteresting_dimensions(data):
    """
    Computes the set of dimensions that only ever take on a
    single value across the entire dataset.
    """
    values_per_dimension = defaultdict(set)

    for dimension_set in data:
        for dimension, value in dimension_set:
            values_per_dimension[dimension].add(value)

    uninteresting = set()

    for dimension, distinct_values in values_per_dimension.items():
        if len(distinct_values) == 1:
            uninteresting.add(dimension)

    return uninteresting


def collect_data(
    tag: str, binary_dir: Path, artifacts: Optional[Path], pytest_opts: str
):
    """
    Executes the specified test using the provided firecracker binaries and
    stores results into the `test_results/tag` directory
    """
    binary_dir = binary_dir.resolve()

    print(
        f"Collecting samples | binaries path: {binary_dir}"
        + f" | artifacts path: {artifacts}"
        if artifacts
        else ""
    )
    test_path = f"test_results/{tag}"
    test_report_path = f"{test_path}/test-report.json"

    # It is not possible to just download them here this script is usually run inside docker
    # and artifacts downloading does not work inside it.
    if artifacts:
        subprocess.run(
            f"./tools/devtool set_current_artifacts {artifacts}", check=True, shell=True
        )

    subprocess.run(
        f"./tools/test.sh --binary-dir={binary_dir} {pytest_opts} -m '' --json-report-file=../{test_report_path}",
        env=os.environ,
        check=True,
        shell=True,
    )

    return load_data_series(Path(test_path))


def check_regression(
    a_samples: List[float], b_samples: List[float], *, n_resamples: int = 9999
):
    """Checks for a regression by performing a permutation test. A permutation test is a non-parametric test that takes
    three parameters: Two populations (sets of samples) and a function computing a "statistic" based on two populations.
    First, the test computes the statistic for the initial populations. It then randomly
    permutes the two populations (e.g. merges them and then randomly splits them again). For each such permuted
    population, the statistic is computed. Then, all the statistics are sorted, and the percentile of the statistic for the
    initial populations is computed. We then look at the fraction of statistics that are larger/smaller than that of the
    initial populations. The minimum of these two fractions will then become the p-value.

    The idea is that if the two populations are indeed drawn from the same distribution (e.g. if performance did not
    change), then permuting will not affect the statistic (indeed, it should be approximately normal-distributed, and
    the statistic for the initial populations will be somewhere "in the middle").

    Useful for performance tests.
    """
    return scipy.stats.permutation_test(
        (a_samples, b_samples),
        # Compute the difference of means, such that a positive different indicates potential for regression.
        lambda x, y: statistics.mean(y) - statistics.mean(x),
        vectorized=False,
        n_resamples=n_resamples,
    )


def analyze_data(
    data_a,
    data_b,
    p_thresh,
    strength_abs_thresh,
    noise_threshold,
    *,
    n_resamples: int = 9999,
):
    """
    Analyzes the A/B-test data produced by `collect_data`, by performing regression tests
    as described this script's doc-comment.

    Returns a mapping of dimensions and properties/metrics to the result of their regression test.
    """
    assert set(data_a.keys()) == set(
        data_b.keys()
    ), "A and B run produced incomparable data. This is a bug in the test!"

    results = {}

    for dimension_set in data_a:
        metrics_a = data_a[dimension_set]
        metrics_b = data_b[dimension_set]

        assert set(metrics_a.keys()) == set(
            metrics_b.keys()
        ), "A and B run produced incomparable data. This is a bug in the test!"

        for metric, (values_a, unit) in metrics_a.items():
            result = check_regression(
                values_a, metrics_b[metric][0], n_resamples=n_resamples
            )
            results[dimension_set, metric] = (result, unit)

    # We sort our A/B-Testing results keyed by metric here. The resulting lists of values
    # will be approximately normal distributed, and we will use this property as a means of error correction.
    # The idea behind this is that testing the same metric (say, restore_latency) across different scenarios (e.g.
    # different vcpu counts) will be related in some unknown way (meaning most scenarios will show a change in the same
    # direction). In particular, if one scenario yields a slight improvement and the next yields a
    # slight degradation, we take this as evidence towards both being mere noise that cancels out.
    #
    # Empirical evidence for this assumption is that
    #  1. Historically, a true performance change has never shown up in just a single test, it always showed up
    #     across most (if not all) tests for a specific metric.
    #  2. Analyzing data collected from historical runs shows that across different parameterizations of the same
    #     metric, the collected samples approximately follow mean / variance = const, with the constant independent
    #     of the parameterization.
    #
    # Mathematically, this has the following justification: By the central
    # limit theorem, the means of samples are (approximately) normal distributed. Denote by A
    # and B the distributions of the mean of samples from the 'A' and 'B'
    # tests respectively. Under our null hypothesis, the distributions of the
    # 'A' and 'B' samples are identical (although we dont know what the exact
    # distributions are), meaning so are A and B, say A ~ B ~ N(mu, sigma^2).
    # The difference of two normal distributions is also normal distributed,
    # with the means being subtracted and the variances being added.
    # Therefore, A - B ~ N(0, 2sigma^2). If we now normalize this distribution by mu (which
    # corresponds to considering the distribution of relative regressions instead), we get (A-B)/mu ~ N(0, c), with c
    # being the constant from point 2. above. This means that we can combine the relative means across
    # different parameterizations, and get a distributions whose expected
    # value is 0, provided our null hypothesis was true. It is exactly this distribution
    # for which we collect samples in the dictionary below. Therefore, a sanity check
    # on the average of the average of the performance changes for a single metric
    # is a good candidates for a sanity check against false-positives.
    #
    # Note that with this approach, for performance changes to "cancel out", we would need essentially a perfect split
    # between scenarios that improve performance and scenarios that degrade performance, something we have not
    # ever observed to actually happen.
    relative_changes_by_metric = defaultdict(list)
    relative_changes_significant = defaultdict(list)

    failures = []
    for (dimension_set, metric), (result, unit) in results.items():
        if is_ignored(dict(dimension_set) | {"metric": metric}):
            continue

        print(f"Doing A/B-test for dimensions {dimension_set} and property {metric}")

        values_a = data_a[dimension_set][metric][0]
        baseline_mean = statistics.mean(values_a)

        relative_changes_by_metric[metric].append(result.statistic / baseline_mean)

        if result.pvalue < p_thresh and abs(result.statistic) > strength_abs_thresh:
            failures.append((dimension_set, metric, result, unit))

            relative_changes_significant[metric].append(
                result.statistic / baseline_mean
            )

    messages = []
    do_not_print_list = uninteresting_dimensions(data_a)
    for dimension_set, metric, result, unit in failures:
        # Sanity check as described above
        if abs(statistics.mean(relative_changes_by_metric[metric])) <= noise_threshold:
            continue

        # No data points for this metric were deemed significant
        if metric not in relative_changes_significant:
            continue

        # The significant data points themselves are above the noise threshold
        if abs(statistics.mean(relative_changes_significant[metric])) > noise_threshold:
            old_mean = statistics.mean(data_a[dimension_set][metric][0])
            new_mean = statistics.mean(data_b[dimension_set][metric][0])

            msg = (
                f"\033[0;32m[Firecracker A/B-Test Runner]\033[0m A/B-testing shows a change of "
                f"{format_with_reduced_unit(result.statistic, unit)}, or {result.statistic / old_mean:.2%}, "
                f"(from {format_with_reduced_unit(old_mean, unit)} to {format_with_reduced_unit(new_mean, unit)}) "
                f"for metric \033[1m{metric}\033[0m with \033[0;31m\033[1mp={result.pvalue}\033[0m. "
                f"This means that observing a change of this magnitude or worse, assuming that performance "
                f"characteristics did not change across the tested commits, has a probability of {result.pvalue:.2%}. "
                f"Tested Dimensions:\n{json.dumps({k: v for k, v in dimension_set if k not in do_not_print_list}, indent=2, sort_keys=True)}"
            )
            messages.append(msg)

    assert not messages, "\n" + "\n".join(messages)
    print("No regressions detected!")


T = TypeVar("T")
U = TypeVar("U")


def binary_ab_test(
    test_runner: Callable[[Path, Optional[Path], bool], T],
    comparator: Callable[[T, T], U],
    *,
    a_directory: Path,
    b_directory: Path,
    a_artifacts: Optional[Path],
    b_artifacts: Optional[Path],
):
    """
    Similar to `git_ab_test`, but instead of locally checking out different revisions, it operates on
    directories containing firecracker/jailer binaries
    """
    result_a = test_runner(a_directory, a_artifacts, True)
    result_b = test_runner(b_directory, b_artifacts, False)

    return result_a, result_b, comparator(result_a, result_b)


def ab_performance_test(
    a_directory: Path,
    b_directory: Path,
    a_artifacts: Optional[Path],
    b_artifacts: Optional[Path],
    pytest_opts,
    p_thresh,
    strength_abs_thresh,
    noise_threshold,
):
    """Does an A/B-test of the specified test with the given firecracker/jailer binaries"""

    return binary_ab_test(
        lambda bin_dir, art_dir, is_a: collect_data(
            is_a and "A" or "B", bin_dir, art_dir, pytest_opts
        ),
        lambda ah, be: analyze_data(
            ah,
            be,
            p_thresh,
            strength_abs_thresh,
            noise_threshold,
            n_resamples=int(100 / p_thresh),
        ),
        a_directory=a_directory,
        b_directory=b_directory,
        a_artifacts=a_artifacts,
        b_artifacts=b_artifacts,
    )


if __name__ == "__main__":
    parser = argparse.ArgumentParser(
        description="Executes Firecracker's A/B testsuite across the specified commits"
    )
    subparsers = parser.add_subparsers(help="commands", dest="command", required=True)
    run_parser = subparsers.add_parser(
        "run",
        help="Run an specific test of our test suite as an A/B-test across two specified commits",
    )
    run_parser.add_argument(
        "--binaries-a",
        help="Directory containing firecracker and jailer binaries to be considered the performance baseline",
        type=Path,
        required=True,
    )
    run_parser.add_argument(
        "--binaries-b",
        help="Directory containing firecracker and jailer binaries whose performance we want to compare against the results from binaries-a",
        type=Path,
        required=True,
    )
    run_parser.add_argument(
        "--artifacts-a",
        help="Name of the artifacts directory in the build/artifacts to use for revision A test. If the directory does not exist, the name will be treated as S3 path and artifacts will be downloaded from there.",
        # Type is string since it can be an s3 path which if passed to `Path` constructor
        # will be incorrectly modified
        type=str,
        required=False,
    )
    run_parser.add_argument(
        "--artifacts-b",
        help="Name of the artifacts directory in the build/artifacts to use for revision B test. If the directory does not exist, the name will be treated as S3 path and artifacts will be downloaded from there.",
        # Type is string since it can be an s3 path which if passed to `Path` constructor
        # will be incorrectly modified
        type=str,
        required=False,
    )
    run_parser.add_argument(
        "--pytest-opts",
        help="Parameters to pass through to pytest, for example for test selection",
        required=True,
    )
    analyze_parser = subparsers.add_parser(
        "analyze",
        help="Analyze the results of two manually ran tests based on their test-report.json files",
    )
    analyze_parser.add_argument(
        "path_a",
        help="The path to the directory with A run",
        type=Path,
    )
    analyze_parser.add_argument(
        "path_b",
        help="The path to the directory with B run",
        type=Path,
    )
    parser.add_argument(
        "--significance",
        help="The p-value threshold that needs to be crossed for a test result to be considered significant",
        type=float,
        default=0.01,
    )
    parser.add_argument(
        "--absolute-strength",
        help="The minimum absolute delta required before a regression will be considered valid",
        type=float,
        default=0.0,
    )
    parser.add_argument(
        "--noise-threshold",
        help="The minimal delta which a metric has to regress on average across all tests that emit it before the regressions will be considered valid.",
        type=float,
        default=0.05,
    )
    args = parser.parse_args()

    if args.command == "run":
        ab_performance_test(
            args.binaries_a,
            args.binaries_b,
            args.artifacts_a,
            args.artifacts_b,
            args.pytest_opts,
            args.significance,
            args.absolute_strength,
            args.noise_threshold,
        )
    else:
        data_a = load_data_series(args.path_a)
        data_b = load_data_series(args.path_b)

        analyze_data(
            data_a,
            data_b,
            args.significance,
            args.absolute_strength,
            args.noise_threshold,
        )