Join bytes in protocols instead of concatenating

[mabdullahsoyturk]

In h11 and httptools, uvicorn concatenates the coming bytes into a bytes attribute. This incurs unnecessary overhead, especially for big files. Here is a benchmark to simulate the behavior:

import os
import subprocess
import sys

BENCHMARK_RUNS = 100
BYTES_SIZE_1 = 0    # The size of the first byte string.
BYTES_SIZE_2 = 4096 # The size of the second byte string to concatenate.


# Template for the 'bytes1 += bytes2' operation.
plus_equals_template = f"""
import os

# These variables are passed in from the main script's configuration.
bytes_size_1 = {BYTES_SIZE_1}
bytes_size_2 = {BYTES_SIZE_2}

# Setup the byte strings
bytes1 = os.urandom(bytes_size_1)
bytes2 = os.urandom(bytes_size_2)

# Perform the operation to be benchmarked
bytes1 += bytes2
"""

# Template for the 'b"".join([bytes1, bytes2])' operation.
join_template = f"""
import os

# These variables are passed in from the main script's configuration.
bytes_size_1 = {BYTES_SIZE_1}
bytes_size_2 = {BYTES_SIZE_2}

# Setup the byte strings
bytes1 = os.urandom(bytes_size_1)
bytes2 = os.urandom(bytes_size_2)

# Perform the operation to be benchmarked
result = b"".join([bytes1, bytes2])
"""

if __name__ == "__main__":
    # Define file names for the temporary scripts
    plus_equals_filename = "temp_benchmark_bytes_plus_equals.py"
    join_filename = "temp_benchmark_bytes_join.py"

    # Write the templates to the temporary files
    with open(plus_equals_filename, "w") as f:
        f.write(plus_equals_template)
    with open(join_filename, "w") as f:
        f.write(join_template)

    print(f"Running each command {BENCHMARK_RUNS} times...")
    print(f"Byte string sizes: {BYTES_SIZE_1 / (1024):.1f} KB and {BYTES_SIZE_2 / (1024):.1f} KB\n")

    # Construct the hyperfine command
    command = [
        "hyperfine",
        "--runs", str(BENCHMARK_RUNS),
        "--warmup", "3",
        f"python {plus_equals_filename}",
        f"python {join_filename}"
    ]

    try:
        result = subprocess.run(command, check=True)

    except FileNotFoundError:
        print("Please install hyperfine to run this benchmark.")
        sys.exit(1)
    except subprocess.CalledProcessError as e:
        print(f"ERROR: Hyperfine exited with an error (code {e.returncode}).")
        sys.exit(1)

On my machine, joining instead of concatenating improves the performance ~17-20%.

So, I suggest to use joining instead of concatenating. If approved, I can prepare the PR.

[Kludex]

Maybe we can use bytearray?

I would like to have a benchmark as part of the test suite using https://codspeed.io/.

I think a good start would be to use the TechEmpower benchmark.

---

Would you like to help here?

[mabdullahsoyturk]

My local experiments show that joining actually performs slightly better than bytearray concat but we can dig in a bit more for various cases. I would be happy to help but I don't think there is a test for file upload speed in TechEmpower benchmark.

[Kludex]

We can add it.