How to benchmark Python code with pytest-benchmark

Everett Pompeii

Everett Pompeii

What is Benchmarking?

Benchmarking is the practice of testing the performance of your code to see how fast (latency) or how much work (throughput) it can do. This often overlooked step in software development is crucial for creating and maintaining fast and performant code. Benchmarking provides the necessary metrics for developers to understand how well their code performs under various workloads and conditions. For the same reasons that you write unit and integration tests to prevent feature regressions, you should write benchmarks to prevent performance regressions. Performance bugs are bugs!

Write FizzBuzz in Python

In order to write benchmarks, we need some source code to benchmark. To start off we are going to write a very simple program, FizzBuzz.

The rules for FizzBuzz are as follows:

Write a program that prints the integers from 1 to 100 (inclusive):

  • For multiples of three, print Fizz
  • For multiples of five, print Buzz
  • For multiples of both three and five, print FizzBuzz
  • For all others, print the number

There are many ways to write FizzBuzz. So we’ll go with the my favorite:

for i in range(1, 101):
    if n % 15 == 0:
        print("FizzBuzz")
    elif n % 3 == 0:
        print("Fizz")
    elif n % 5 == 0:
        print("Buzz")
    else:
        print(i)
  • Iterate from 1 to 100, using a range of 101.
  • For each number, calculate the modulus (remainder after division).
  • If the remainder is 0, then the number is a multiple of the given factor:
    • If the remainder is 0 for 15, then print FizzBuzz.
    • If the remainder is 0 for 3, then print Fizz.
    • If the remainder is 0 for 5 ,then print Buzz.
  • Otherwise, just print the number.

Follow Step-by-Step

In order to follow along with this set-by-step tutorial, you will need to install Python and install pipenv.

🐰 The source code for this post is available on GitHub.

Create a Python file named game.py, and set its contents to the above FizzBuzz implementation.

Then run python game.py. The output should look like:

$ python game.py
1
2
Fizz
4
Buzz
Fizz
7
8
Fizz
Buzz
11
Fizz
13
14
FizzBuzz
...
97
98
Fizz
Buzz

🐰 Boom! You’re cracking the coding interview!

Before going any further, it is important to discuss the differences between micro-benchmarking and macro-benchmarking.

Micro-Benchmarking vs Macro-Benchmarking

There are two major categories of software benchmarks: micro-benchmarks and macro-benchmarks. Micro-benchmarks operate at a level similar to unit tests. For example, a benchmark for a function that determines Fizz, Buzz, or FizzBuzz for a single number would be a micro-benchmark. Macro-benchmarks operate at a level similar to integration tests. For example, a benchmark for a function that plays the entire game of FizzBuzz, from 1 to 100, would be a macro-benchmark.

Generally, it is best to test at the lowest level of abstraction possible. In the case benchmarks, this makes them both easier to maintain, and it helps to reduce the amount of noise in the measurements. However, just as having some end-to-end tests can be very useful for sanity checking the entire system comes together as expected, having macro-benchmarks can be very useful for making sure that the critical paths through your software remain performant.

Benchmarking in Python

The two popular options for benchmarking in Python are: pytest-benchmark and airspeed velocity (asv).

pytest-benchmark is a powerful benchmarking tool integrated with the popular pytest testing framework. It allows developers to measure and compare the performance of their code by running benchmarks alongside their unit tests. Users can easily compare their benchmark results locally and export their results in various formats, such as JSON.

airspeed velocity (asv) is another advanced benchmarking tool in the Python ecosystem. One of the key benefits of asv is its ability to generate detailed and interactive HTML reports, which make it easy to visualize performance trends and identify regressions. Additionally, asv supports Relative Continuous Benchmarking out of the box.

Both are support by Bencher. So why choose pytest-benchmark? pytest-benchmark integrates seamlessly with pytest, which is the de facto standard unit test harness in the Python ecosystem. I would suggest using pytest-benchmark for benchmarking your code’s latency, especially if you are already using pytest. That is, pytest-benchmark is great for measuring wall clock time.

Refactor FizzBuzz

In order to test our FizzBuzz application, we need to decouple our logic from our program’s main execution. Benchmark harnesses can’t benchmark the main execution. In order to do this, we need to make few changes.

Let’s refactor our FizzBuzz logic into a couple of function:

def play_game(n, should_print):
    result = fizz_buzz(n)
    if should_print:
        print(result)
    return result


def fizz_buzz(n):
    if n % 15 == 0:
       return "FizzBuzz"
    elif n % 3 == 0:
        return "Fizz"
    elif n % 5 == 0:
        return "Buzz"
    else:
        return str(n)
  • play_game: Takes in an integer n, calls fizz_buzz with that number, and if should_print is True print the result.
  • fizz_buzz: Takes in an integer n and performs the actual Fizz, Buzz, FizzBuzz, or number logic returning the result as a string.

Then update then main execution to look like this:

game.py
for i in range(1, 101):
    play_game(i, True)

The main execution for our program iterates through the numbers 1 to 100 inclusive and calls play_game for each number, with should_print set to True.

Benchmarking FizzBuzz

In order to benchmark our code, we need to create a test function that runs our benchmark. At the bottom of game.py add the following code:

def test_game(benchmark):
    def run_game():
        for i in range(1, 101):
            play_game(i, False)
    benchmark(run_game)
  • Create a function named test_game that takes in a pytest-benchmark benchmark fixture.
  • Create a run_game function that iterate from 1 to 100 inclusively.
    • For each number, call play_game, with should_print set to False.
  • Pass the run_game function the the benchmark runner.

Now we need to configure our project to run our benchmarks.

Create a new virtual environment with pipenv:

$ pipenv shell


Creating a Pipfile for this project...
Launching subshell in virtual environment...
 source /usr/bencher/.local/share/virtualenvs/test-xnizGmtA/bin/activate

Install pytest-benchmark inside of that new pipenv environment:

$ pipenv install pytest-benchmark
Creating a Pipfile for this project...
Installing pytest-benchmark...
Resolving pytest-benchmark...
Added pytest-benchmark to Pipfile's [packages] ...
✔ Installation Succeeded
Pipfile.lock not found, creating...
Locking [packages] dependencies...
Building requirements...
Resolving dependencies...
✔ Success!
Locking [dev-packages] dependencies...
Updated Pipfile.lock (be953321071292b6175f231c7e2e835a3cd26169a0d52b7b781b344d65e8cce3)!
Installing dependencies from Pipfile.lock (e8cce3)...

Now we’re ready to benchmark our code, run pytest game.py:

$ pytest game.py
======================================================= test session starts ========================================================
platform darwin -- Python 3.12.7, pytest-8.3.3, pluggy-1.5.0
benchmark: 4.0.0 (defaults: timer=time.perf_counter disable_gc=False min_rounds=5 min_time=0.000005 max_time=1.0 calibration_precision=10 warmup=False warmup_iterations=100000)
rootdir: /usr/bencher/examples/python/pytest_benchmark
plugins: benchmark-4.0.0
collected 1 item


game.py .                                                                                                                    [100%]




------------------------------------------------- benchmark: 1 tests -------------------------------------------------
Name (time in us)         Min       Max     Mean  StdDev   Median     IQR   Outliers  OPS (Kops/s)  Rounds  Iterations
----------------------------------------------------------------------------------------------------------------------
test_game             10.5416  237.7499  10.8307  1.3958  10.7088  0.1248  191;10096       92.3304   57280           1
----------------------------------------------------------------------------------------------------------------------


Legend:
  Outliers: 1 Standard Deviation from Mean; 1.5 IQR (InterQuartile Range) from 1st Quartile and 3rd Quartile.
  OPS: Operations Per Second, computed as 1 / Mean
======================================================== 1 passed in 1.68s =========================================================

🐰 Lettuce turnip the beet! We’ve got our first benchmark metrics!

Finally, we can rest our weary developer heads… Just kidding, our users want a new feature!

Write FizzBuzzFibonacci in Python

Our Key Performance Indicators (KPIs) are down, so our Product Manager (PM) wants us to add a new feature. After much brainstorming and many user interviews, it is decided that good ole FizzBuzz isn’t enough. Kids these days want a new game, FizzBuzzFibonacci.

The rules for FizzBuzzFibonacci are as follows:

Write a program that prints the integers from 1 to 100 (inclusive):

  • For multiples of three, print Fizz
  • For multiples of five, print Buzz
  • For multiples of both three and five, print FizzBuzz
  • For numbers that are part of the Fibonacci sequence, only print Fibonacci
  • For all others, print the number

The Fibonacci sequence is a sequence in which each number is the sum of the two preceding numbers. For example, starting at 0 and 1 the next number in the Fibonacci sequence would be 1. Followed by: 2, 3, 5, 8 and so on. Numbers that are part of the Fibonacci sequence are known as Fibonacci numbers. So we’re going to have to write a function that detects Fibonacci numbers.

There are many ways to write the Fibonacci sequence and likewise many ways to detect a Fibonacci number. So we’ll go with the my favorite:

def is_fibonacci_number(n):
    for i in range(n + 1):
        previous, current = 0, 1
        while current < i:
            next_value = previous + current
            previous = current
            current = next_value
        if current == n:
            return True
    return False
  • Create a function named is_fibonacci_number that takes in an integer and returns a boolean.
  • Iterate for all number from 0 to our given number n inclusive.
  • Initialize our Fibonacci sequence starting with 0 and 1 as the previous and current numbers respectively.
  • Iterate while the current number is less than the current iteration i.
  • Add the previous and current number to get the next_value number.
  • Update the previous number to the current number.
  • Update the current number to the next_value number.
  • Once current is greater than or equal to the given number n, we will exit the loop.
  • Check to see is the current number is equal to the given number n and if so return True.
  • Otherwise, return False.

Now we will need to update our fizz_buzz function:

def fizz_buzz_fibonacci(n):
    if is_fibonacci_number(n):
        return "Fibonacci"
    elif n % 15 == 0:
        return "FizzBuzz"
    elif n % 3 == 0:
        return "Fizz"
    elif n % 5 == 0:
        return "Buzz"
    else:
        return str(n)
  • Rename the fizz_buzz function to fizz_buzz_fibonacci to make it more descriptive.
  • Call our is_fibonacci_number helper function.
  • If the result from is_fibonacci_number is True then return Fibonacci.
  • If the result from is_fibonacci_number is False then perform the same Fizz, Buzz, FizzBuzz, or number logic returning the result.

Because we renamed fizz_buzz to fizz_buzz_fibonacci we also need to update our play_game function:

def play_game(n, should_print):
    result = fizz_buzz_fibonacci(n)
    if should_print:
        print(result)
    return result

Both our main execution and the test_game function can stay exactly the same.

Benchmarking FizzBuzzFibonacci

Now we can rerun our benchmark:

$ pytest game.py
======================================================= test session starts ========================================================
platform darwin -- Python 3.12.7, pytest-8.3.3, pluggy-1.5.0
benchmark: 4.0.0 (defaults: timer=time.perf_counter disable_gc=False min_rounds=5 min_time=0.000005 max_time=1.0 calibration_precision=10 warmup=False warmup_iterations=100000)
rootdir: /usr/bencher/examples/python/pytest_benchmark
plugins: benchmark-4.0.0
collected 1 item


game.py .                                                                                                                    [100%]




--------------------------------------------------- benchmark: 1 tests --------------------------------------------------
Name (time in us)          Min       Max      Mean   StdDev    Median     IQR  Outliers  OPS (Kops/s)  Rounds  Iterations
-------------------------------------------------------------------------------------------------------------------------
test_game             726.9592  848.2919  735.5682  13.4925  731.4999  4.7078   146;192        1.3595    1299           1
-------------------------------------------------------------------------------------------------------------------------


Legend:
  Outliers: 1 Standard Deviation from Mean; 1.5 IQR (InterQuartile Range) from 1st Quartile and 3rd Quartile.
  OPS: Operations Per Second, computed as 1 / Mean
======================================================== 1 passed in 1.97s =========================================================

Scrolling back through our terminal history, we can make an eyeball comparison between the performance of our FizzBuzz and FizzBuzzFibonacci games: 10.8307 us vs 735.5682 us. Your numbers will be a little different than mine. However, the difference between the two games is likely in the 50x range. That seems good to me! Especially for adding a feature as fancy sounding as Fibonacci to our game. The kids will love it!

Expand FizzBuzzFibonacci in Python

Our game is a hit! The kids do indeed love playing FizzBuzzFibonacci. So much so that word has come down from the execs that they want a sequel. But this is the modern world, we need Annual Recurring Revenue (ARR) not one time purchases! The new vision for our game is that it is open ended, no more living between the bounds of 1 and 100 (even if they are inclusive). No, we’re on to new frontiers!

The rules for Open World FizzBuzzFibonacci are as follows:

Write a program that takes in any positive integer and prints:

  • For multiples of three, print Fizz
  • For multiples of five, print Buzz
  • For multiples of both three and five, print FizzBuzz
  • For numbers that are part of the Fibonacci sequence, only print Fibonacci
  • For all others, print the number

In order to have our game work for any number, we will need to accept a command line argument. Update the main execution to look like this:

game.py
import sys


args = sys.argv
if len(args) > 1 and args[1].isdigit():
    i = int(args[1])
    play_game(i, True)
else:
    print("Please, enter a positive integer to play...")
  • Import the sys package.
  • Collect all of the arguments (args) passed to our game from the command line.
  • Get the first argument passed to our game and check to see if it is a digit.
    • If so, parse the first argument as an integer, i.
    • Play our game with the newly parsed integer i.
  • If parsing fails or no argument is passed in, default to prompting for a valid input.

Now we can play our game with any number! Run python game.py followed by an integer to play our game:

$ python game.py 9
Fizz
$ python game.py 10
Buzz
$ python game.py 13
Fibonacci

And if we omit or provide an invalid number:

$ python game.py
Please, enter a positive integer to play...
$ python game.py bad
Please, enter a positive integer to play...

Wow, that was some thorough testing! CI passes. Our bosses are thrilled. Let’s ship it! 🚀

The End


SpongeBob SquarePants Three Weeks Later
This is Fine meme

🐰 … the end of your career maybe?


Just kidding! Everything is on fire! 🔥

Well, at first everything seemed to be going fine. And then at 02:07 AM on Saturday my pager went off:

📟 Your game is on fire! 🔥

After scrambling out of bed, I tried to figure out what was going on. I tried to search through the logs, but that was hard because everything kept crashing. Finally, I found the issue. The kids! They loved our game so much, they were playing it all the way up to a million! In a flash of brilliance, I added two new benchmarks:

def test_game_100(benchmark):
    def run_game():
        play_game(100, False)
    benchmark(run_game)


def test_game_1_000_000(benchmark):
    def run_game():
        play_game(1_000_000, False)
    benchmark(run_game)
  • A micro-benchmark test_game_100 for playing the game with the number one hundred (100)
  • A micro-benchmark test_game_1_000_000 for playing the game with the number one million (1_000_000)

When I ran it, I got this:

$ pytest game.py
======================================================= test session starts ========================================================
platform darwin -- Python 3.12.7, pytest-8.3.3, pluggy-1.5.0
benchmark: 4.0.0 (defaults: timer=time.perf_counter disable_gc=False min_rounds=5 min_time=0.000005 max_time=1.0 calibration_precision=10 warmup=False warmup_iterations=100000)
rootdir: /usr/bencher/examples/python/pytest_benchmark
plugins: benchmark-4.0.0
collected 3 items


game.py ...                                                                                                                  [100%]

Wait for it… wait for it…

-------------------------------------------------------------------------------------------------- benchmark: 3 tests --------------------------------------------------------------------------------------------------
Name (time in us)                Min                     Max                    Mean                StdDev                  Median                    IQR            Outliers          OPS            Rounds  Iterations
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
test_game_100                15.4166 (1.0)          112.8749 (1.0)           15.8470 (1.0)          1.1725 (1.0)           15.6672 (1.0)           0.1672 (1.0)     1276;7201  63,103.3078 (1.0)       58970           1
test_game                   727.0002 (47.16)      1,074.3327 (9.52)         754.3231 (47.60)       33.2047 (28.32)        748.9999 (47.81)        33.7283 (201.76)     134;54   1,325.6918 (0.02)       1319           1
test_game_1_000_000     565,232.3328 (>1000.0)  579,829.1252 (>1000.0)  571,684.6334 (>1000.0)  6,365.1577 (>1000.0)  568,294.3747 (>1000.0)  10,454.0113 (>1000.0)       2;0       1.7492 (0.00)          5           1
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------


Legend:
  Outliers: 1 Standard Deviation from Mean; 1.5 IQR (InterQuartile Range) from 1st Quartile and 3rd Quartile.
  OPS: Operations Per Second, computed as 1 / Mean
======================================================== 3 passed in 7.01s =========================================================

What! 15.8470 us x 1,000 should be 15,847.0 us not 571,684.6334 us 🤯 Even though I got my Fibonacci sequence code functionally correct, I must have a performance bug in there somewhere.

Fix FizzBuzzFibonacci in Python

Let’s take another look at that is_fibonacci_number function:

def is_fibonacci_number(n):
    for i in range(n + 1):
        previous, current = 0, 1
        while current < i:
            next_value = previous + current
            previous = current
            current = next_value
        if current == n:
            return True
    return False

Now that I’m thinking about performance, I do realize that I have an unnecessary, extra loop. We can completely get rid of the for i in range(n + 1): loop and just compare the current value to the given number (n) 🤦

def is_fibonacci_number(n):
    previous, current = 0, 1
    while current < n:
        next_value = previous + current
        previous = current
        current = next_value
    return current == n
  • Update our is_fibonacci_number function.
  • Initialize our Fibonacci sequence starting with 0 and 1 as the previous and current numbers respectively.
  • Iterate while the current number is less than the given number n.
  • Add the previous and current number to get the next_value number.
  • Update the previous number to the current number.
  • Update the current number to the next_value number.
  • Once current is greater than or equal to the given number n, we will exit the loop.
  • Check to see if the current number is equal to the given number n and return that result.

Now lets rerun those benchmarks and see how we did:

$ pytest game.py
======================================================= test session starts ========================================================
platform darwin -- Python 3.12.7, pytest-8.3.3, pluggy-1.5.0
benchmark: 4.0.0 (defaults: timer=time.perf_counter disable_gc=False min_rounds=5 min_time=0.000005 max_time=1.0 calibration_precision=10 warmup=False warmup_iterations=100000)
rootdir: /usr/bencher/examples/python/pytest_benchmark
plugins: benchmark-4.0.0
collected 3 items


game.py ...                                                                                                                  [100%]




------------------------------------------------------------------------------------------------ benchmark: 3 tests ------------------------------------------------------------------------------------------------
Name (time in ns)               Min                     Max                   Mean                StdDev                 Median                 IQR             Outliers  OPS (Kops/s)            Rounds  Iterations
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
test_game_100              309.8685 (1.0)       40,197.8614 (2.38)        322.0815 (1.0)        101.7570 (1.0)         320.2877 (1.0)        5.1805 (1.0)      321;12616    3,104.8046 (1.0)      195120          16
test_game_1_000_000        724.9881 (2.34)      16,912.4920 (1.0)         753.1445 (2.34)       121.0458 (1.19)        741.7053 (2.32)      12.4797 (2.41)     656;13698    1,327.7664 (0.43)     123073          10
test_game               26,958.9946 (87.00)    129,667.1107 (7.67)     27,448.7719 (85.22)    1,555.0003 (15.28)    27,291.9424 (85.21)    165.7754 (32.00)     479;2372       36.4315 (0.01)      25918           1
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------


Legend:
  Outliers: 1 Standard Deviation from Mean; 1.5 IQR (InterQuartile Range) from 1st Quartile and 3rd Quartile.
  OPS: Operations Per Second, computed as 1 / Mean
======================================================== 3 passed in 3.99s =========================================================

Oh, wow! Our test_game benchmark is back down to around where it was for the original FizzBuzz. I wish I could remember exactly what that score was. It’s been three weeks though. My terminal history doesn’t go back that far. And pytest-benchmark only stores its results when we ask it to. But I think it’s close!

The test_game_100 benchmark is down nearly 50x to 322.0815 ns. And the test_game_1_000_000 benchmark is down more than 500,000x! 571,684,633.4 ns to 753.1445 ns!

🐰 Hey, at least we caught this performance bug before it made it to production… oh, right. Nevermind…

Catch Performance Regressions in CI

The execs weren’t happy about the deluge of negative reviews our game received due to my little performance bug. They told me not to let it happen again, and when I asked how, they just told me not to do it again. How am I supposed to manage that‽

Luckily, I’ve found this awesome open source tool called Bencher. There’s a super generous free tier, so I can just use Bencher Cloud for my personal projects. And at work where everything needs to be in our private cloud, I’ve started using Bencher Self-Hosted.

Bencher has a built-in adapters, so it’s easy to integrate into CI. After following the Quick Start guide, I’m able to run my benchmarks and track them with Bencher.

$ bencher run --adapter python_pytest --file results.json "pytest --benchmark-json results.json game.py"
======================================================= test session starts ========================================================
platform darwin -- Python 3.12.7, pytest-8.3.3, pluggy-1.5.0
benchmark: 4.0.0 (defaults: timer=time.perf_counter disable_gc=False min_rounds=5 min_time=0.000005 max_time=1.0 calibration_precision=10 warmup=False warmup_iterations=100000)
rootdir: /usr/bencher/examples/python/pytest_benchmark
plugins: benchmark-4.0.0
collected 3 items


game.py ...


...


View results:
- test_game (Latency): https://bencher.dev/console/projects/game/perf?measures=52507e04-ffd9-4021-b141-7d4b9f1e9194&branches=3a27b3ce-225c-4076-af7c-75adbc34ef9a&testbeds=bc05ed88-74c1-430d-b96a-5394fdd18bb0&benchmarks=077449e5-5b45-4c00-bdfb-3a277413180d&start_time=1697224006000&end_time=1699816009000&upper_boundary=true
- test_game_100 (Latency): https://bencher.dev/console/projects/game/perf?measures=52507e04-ffd9-4021-b141-7d4b9f1e9194&branches=3a27b3ce-225c-4076-af7c-75adbc34ef9a&testbeds=bc05ed88-74c1-430d-b96a-5394fdd18bb0&benchmarks=96508869-4fa2-44ac-8e60-b635b83a17b7&start_time=1697224006000&end_time=1699816009000&upper_boundary=true
- test_game_1_000_000 (Latency): https://bencher.dev/console/projects/game/perf?measures=52507e04-ffd9-4021-b141-7d4b9f1e9194&branches=3a27b3ce-225c-4076-af7c-75adbc34ef9a&testbeds=bc05ed88-74c1-430d-b96a-5394fdd18bb0&benchmarks=ff014217-4570-42ea-8813-6ed0284500a4&start_time=1697224006000&end_time=1699816009000&upper_boundary=true

Using this nifty time travel device that a nice rabbit gave me, I was able to go back in time and replay what would have happened if we were using Bencher all along. You can see where we first pushed the buggy FizzBuzzFibonacci implementation. I immediately got failures in CI as a comment on my pull request. That same day, I fixed the performance bug, getting rid of that needless, extra loop. No fires. Just happy users.

Bencher: Continuous Benchmarking

🐰 Bencher

Bencher is a suite of continuous benchmarking tools. Have you ever had a performance regression impact your users? Bencher could have prevented that from happening. Bencher allows you to detect and prevent performance regressions before they make it to production.

  • Run: Run your benchmarks locally or in CI using your favorite benchmarking tools. The bencher CLI simply wraps your existing benchmark harness and stores its results.
  • Track: Track the results of your benchmarks over time. Monitor, query, and graph the results using the Bencher web console based on the source branch, testbed, benchmark, and measure.
  • Catch: Catch performance regressions in CI. Bencher uses state of the art, customizable analytics to detect performance regressions before they make it to production.

For the same reasons that unit tests are run in CI to prevent feature regressions, benchmarks should be run in CI with Bencher to prevent performance regressions. Performance bugs are bugs!

Start catching performance regressions in CI — try Bencher Cloud for free.


Published: Wed, October 9, 2024 at 4:30:00 PM UTC | Last Updated: Sun, November 3, 2024 at 8:45:00 AM UTC