Add Starlette: the ASGI framework that powers FastAPI (~12k ⭐)

[BennyFranciscus]

Starlette — The little ASGI framework that shines 🌟

Adds Starlette (~12k ⭐) to HttpArena.

Why Starlette?

FastAPI is already in HttpArena — and FastAPI is literally built on Starlette. Having both answers the question every Python developer asks: how much overhead does FastAPI's Pydantic validation, OpenAPI generation, and dependency injection actually add?

Same ASGI stack (Uvicorn + uvloop + Gunicorn), same runtime, same language — but Starlette is the raw foundation without the convenience layer. This is the purest "framework vs framework-on-framework" comparison possible.

Setup

• Starlette 0.46.1 on Uvicorn (uvloop) + Gunicorn multi-worker
• Pre-computed JSON + gzip caches for /json and /compression
• Thread-local SQLite connections with mmap
• orjson for fast JSON serialization
• Identical worker config to the FastAPI entry for fair comparison

The Python lineup is now:

Entry	Stack	Model
Flask	WSGI/Gunicorn sync	Traditional sync
Starlette	ASGI/Uvicorn async	Raw async foundation
FastAPI	ASGI/Uvicorn async	Starlette + Pydantic + OpenAPI
Django	WSGI/Gunicorn sync	Full-stack ORM framework

cc @tomchristie @alex-grover @aminalaee — would love to see how Starlette stacks up against FastAPI in the benchmarks! The overhead question is one I've been curious about for a while.

[BennyFranciscus]

CI failure is the same port 8080 conflict on the self-hosted runner (stale Docker container) — same issue as Rocket #75 and Koa #77. The build compiles fine and all dependencies install correctly. Should pass on a clean runner.

[BennyFranciscus]

CI is green now ✅ — looks like the port conflict on the runner got resolved. Starlette is ready for benchmarking whenever you want to kick it off!

[MDA2AV]

/benchmark

[github-actions]

🚀 Benchmark run triggered for starlette (all profiles). Results will be posted here when done.

[github-actions]

Benchmark Results

Framework: starlette | Profile: all profiles

starlette / baseline / 512c (p=1, r=0, cpu=unlimited)
  Best: 572933 req/s (CPU: 10610.6%, Mem: 4.9GiB) ===

starlette / baseline / 4096c (p=1, r=0, cpu=unlimited)
  Best: 628030 req/s (CPU: 10497.8%, Mem: 4.9GiB) ===

starlette / baseline / 16384c (p=1, r=0, cpu=unlimited)
  Best: 517806 req/s (CPU: 5797.0%, Mem: 831.5MiB) ===

starlette / pipelined / 512c (p=16, r=0, cpu=unlimited)
  Best: 760652 req/s (CPU: 10316.1%, Mem: 4.8GiB) ===

starlette / pipelined / 4096c (p=16, r=0, cpu=unlimited)
  Best: 799555 req/s (CPU: 10255.6%, Mem: 5.1GiB) ===

starlette / pipelined / 16384c (p=16, r=0, cpu=unlimited)
  Best: 668705 req/s (CPU: 9981.9%, Mem: 5.8GiB) ===

starlette / limited-conn / 512c (p=1, r=10, cpu=unlimited)
  Best: 411518 req/s (CPU: 5745.7%, Mem: 941.6MiB) ===

starlette / limited-conn / 4096c (p=1, r=10, cpu=unlimited)
  Best: 428041 req/s (CPU: 5603.0%, Mem: 950.8MiB) ===

starlette / json / 4096c (p=1, r=0, cpu=unlimited)
  Best: 743130 req/s (CPU: 9989.7%, Mem: 4.9GiB) ===

starlette / json / 16384c (p=1, r=0, cpu=unlimited)
  Best: 604394 req/s (CPU: 9629.4%, Mem: 5.2GiB) ===

starlette / upload / 64c (p=1, r=0, cpu=unlimited)
  Best: 470 req/s (CPU: 5304.4%, Mem: 8.6GiB) ===

starlette / upload / 256c (p=1, r=0, cpu=unlimited)
  Best: 430 req/s (CPU: 8767.8%, Mem: 15.6GiB) ===

starlette / upload / 512c (p=1, r=0, cpu=unlimited)
  Best: 404 req/s (CPU: 2024.3%, Mem: 11.3GiB) ===

starlette / compression / 4096c (p=1, r=0, cpu=unlimited)
  Best: 85395 req/s (CPU: 4304.1%, Mem: 848.2MiB) ===

starlette / compression / 16384c (p=1, r=0, cpu=unlimited)
  Best: 79593 req/s (CPU: 5025.7%, Mem: 758.6MiB) ===

starlette / noisy / 512c (p=1, r=0, cpu=unlimited)
  Best: 474245 req/s (CPU: 10564.8%, Mem: 5.2GiB) ===

starlette / noisy / 4096c (p=1, r=0, cpu=unlimited)
  Best: 534602 req/s (CPU: 10634.7%, Mem: 5.3GiB) ===

starlette / noisy / 16384c (p=1, r=0, cpu=unlimited)
  Best: 455997 req/s (CPU: 6406.1%, Mem: 2.9GiB) ===

starlette / mixed / 4096c (p=1, r=5, cpu=unlimited)
  Best: 70399 req/s (CPU: 5484.1%, Mem: 962.9MiB) ===

starlette / mixed / 16384c (p=1, r=5, cpu=unlimited)
  Best: 62747 req/s (CPU: 4853.6%, Mem: 4.0GiB) ===

Full log

  Status codes: 2xx=1684229, 3xx=0, 4xx=214689, 5xx=0
  Latency samples: 1898911 / 1898918 responses (100.0%)
  Reconnects: 3600
  Errors: connect 0, read 1, timeout 0
  Per-template: 1078586,606051,211005,0,3276
  Per-template-ok: 1078230,605999,0,0,0

  WARNING: 214689/1898918 responses (11.3%) had unexpected status (expected 2xx)
  CPU: 10096.8% | Mem: 6.5GiB

=== Best: 455997 req/s (CPU: 6406.1%, Mem: 2.9GiB) ===
  Input BW: 46.10MB/s (avg template: 106 bytes)
[dry-run] Results not saved (use --save to persist)
httparena-bench-starlette
httparena-bench-starlette

==============================================
=== starlette / mixed / 4096c (p=1, r=5, cpu=unlimited) ===
==============================================
ce21ba9a1b22359f09981bb5540bdac5b63f1b1575b5b5832964d9d2e788b9ae
[wait] Waiting for server...
[ready] Server is up

[run 1/3]
gcannon — io_uring HTTP load generator
  Target:    localhost:8080/
  Threads:   64
  Conns:     4096 (64/thread)
  Pipeline:  1
  Req/conn:  5
  Templates: 10
  Expected:  200
  Duration:  5s


  Thread Stats   Avg      p50      p90      p99    p99.9
    Latency   38.60ms   7.53ms   140.40ms   175.10ms   953.20ms

  384042 requests in 5.01s, 352703 responses
  Throughput: 70.44K req/s
  Bandwidth:  2.79GB/s
  Status codes: 2xx=352703, 3xx=0, 4xx=0, 5xx=0
  Latency samples: 352702 / 352703 responses (100.0%)
  Reconnects: 76271
  Per-template: 38030,38075,38302,38488,38681,30949,31040,38306,30475,30356
  Per-template-ok: 38030,38075,38302,38488,38681,30949,31040,38306,30475,30356
  CPU: 5484.1% | Mem: 962.9MiB

[run 2/3]
gcannon — io_uring HTTP load generator
  Target:    localhost:8080/
  Threads:   64
  Conns:     4096 (64/thread)
  Pipeline:  1
  Req/conn:  5
  Templates: 10
  Expected:  200
  Duration:  5s


  Thread Stats   Avg      p50      p90      p99    p99.9
    Latency   53.21ms   37.80ms   109.60ms   261.60ms   424.70ms

  326075 requests in 5.03s, 296933 responses
  Throughput: 59.02K req/s
  Bandwidth:  2.31GB/s
  Status codes: 2xx=296933, 3xx=0, 4xx=0, 5xx=0
  Latency samples: 296933 / 296933 responses (100.0%)
  Reconnects: 63155
  Per-template: 31294,31634,31985,32469,32875,26783,26994,32566,25317,25016
  Per-template-ok: 31294,31634,31985,32469,32875,26783,26994,32566,25317,25016
  CPU: 8664.1% | Mem: 6.3GiB

[run 3/3]
gcannon — io_uring HTTP load generator
  Target:    localhost:8080/
  Threads:   64
  Conns:     4096 (64/thread)
  Pipeline:  1
  Req/conn:  5
  Templates: 10
  Expected:  200
  Duration:  5s


  Thread Stats   Avg      p50      p90      p99    p99.9
    Latency   69.83ms   60.00ms   127.40ms   223.10ms   310.20ms

  220186 requests in 5.03s, 200044 responses
  Throughput: 39.76K req/s
  Bandwidth:  1.56GB/s
  Status codes: 2xx=200044, 3xx=0, 4xx=0, 5xx=0
  Latency samples: 200044 / 200044 responses (100.0%)
  Reconnects: 42122
  Per-template: 21191,21707,21960,21993,22013,17778,17763,21818,17034,16787
  Per-template-ok: 21191,21707,21960,21993,22013,17778,17763,21818,17034,16787
  CPU: 9585.5% | Mem: 8.0GiB

=== Best: 70399 req/s (CPU: 5484.1%, Mem: 962.9MiB) ===
  Input BW: 6.88GB/s (avg template: 104924 bytes)
[dry-run] Results not saved (use --save to persist)
httparena-bench-starlette
httparena-bench-starlette

==============================================
=== starlette / mixed / 16384c (p=1, r=5, cpu=unlimited) ===
==============================================
14c8bd99681a460ef58b2f2ae2dd9d2a78fd3a7dc1d4f765f7dc09ef2e3bf7e2
[wait] Waiting for server...
[ready] Server is up

[run 1/3]
gcannon — io_uring HTTP load generator
  Target:    localhost:8080/
  Threads:   64
  Conns:     16384 (256/thread)
  Pipeline:  1
  Req/conn:  5
  Templates: 10
  Expected:  200
  Duration:  5s


  Thread Stats   Avg      p50      p90      p99    p99.9
    Latency   159.69ms   78.60ms   209.20ms    3.50s    3.64s

  355776 requests in 5.03s, 315619 responses
  Throughput: 62.73K req/s
  Bandwidth:  2.48GB/s
  Status codes: 2xx=315619, 3xx=0, 4xx=0, 5xx=0
  Latency samples: 315618 / 315619 responses (100.0%)
  Reconnects: 62500
  Errors: connect 0, read 756, timeout 0
  Per-template: 32990,34045,34930,35381,35329,29029,29246,31493,27360,25816
  Per-template-ok: 32990,34045,34930,35381,35329,29029,29246,31493,27360,25816
  CPU: 4853.6% | Mem: 4.0GiB

[run 2/3]
gcannon — io_uring HTTP load generator
  Target:    localhost:8080/
  Threads:   64
  Conns:     16384 (256/thread)
  Pipeline:  1
  Req/conn:  5
  Templates: 10
  Expected:  200
  Duration:  5s


  Thread Stats   Avg      p50      p90      p99    p99.9
    Latency   245.29ms   219.90ms   422.40ms   832.00ms    1.16s

  243400 requests in 5.04s, 214686 responses
  Throughput: 42.60K req/s
  Bandwidth:  1.42GB/s
  Status codes: 2xx=214686, 3xx=0, 4xx=0, 5xx=0
  Latency samples: 214686 / 214686 responses (100.0%)
  Reconnects: 40369
  Per-template: 16961,21107,24467,26659,26843,21942,21903,25689,16914,12201
  Per-template-ok: 16961,21107,24467,26659,26843,21942,21903,25689,16914,12201
  CPU: 7444.4% | Mem: 6.8GiB

[run 3/3]
gcannon — io_uring HTTP load generator
  Target:    localhost:8080/
  Threads:   64
  Conns:     16384 (256/thread)
  Pipeline:  1
  Req/conn:  5
  Templates: 10
  Expected:  200
  Duration:  5s


  Thread Stats   Avg      p50      p90      p99    p99.9
    Latency   300.85ms   274.40ms   522.60ms    1.19s    1.45s

  162573 requests in 5.05s, 140500 responses
  Throughput: 27.82K req/s
  Bandwidth:  842.44MB/s
  Status codes: 2xx=140500, 3xx=0, 4xx=0, 5xx=0
  Latency samples: 140500 / 140500 responses (100.0%)
  Reconnects: 25196
  Per-template: 10725,16035,17234,17296,17185,14308,14187,17105,9096,7329
  Per-template-ok: 10725,16035,17234,17296,17185,14308,14187,17105,9096,7329
  CPU: 8301.5% | Mem: 8.4GiB

=== Best: 62747 req/s (CPU: 4853.6%, Mem: 4.0GiB) ===
  Input BW: 6.13GB/s (avg template: 104924 bytes)
[dry-run] Results not saved (use --save to persist)
httparena-bench-starlette
httparena-bench-starlette
[skip] starlette does not subscribe to baseline-h2
[skip] starlette does not subscribe to static-h2
[skip] starlette does not subscribe to baseline-h3
[skip] starlette does not subscribe to static-h3
[skip] starlette does not subscribe to unary-grpc
[skip] starlette does not subscribe to unary-grpc-tls
[skip] starlette does not subscribe to echo-ws
[restore] Restoring CPU governor to powersave...

[BennyFranciscus]

Starlette benchmarks are in! Really nice numbers for a pure ASGI framework:

Highlights:

• 628K baseline at 4096c — significantly faster than Django (113K) and even beating FastAPI in raw throughput, which makes sense since Starlette is the engine under FastAPI without the extra validation layer
• 800K pipelined at 4096c — solid scaling with connection pipelining
• 743K JSON at 4096c — uvicorn + uvloop doing work here
• 85K compression — reasonable for Python doing gzip on-the-fly
• 470 upload at 64c — similar ballpark to other Python frameworks

The noisy profile at 16384c shows 11.3% unexpected status codes (214K 4xx responses) — worth keeping an eye on, though this might just be uvicorn connection handling under extreme load.

Starlette slots in nicely as the fastest pure-Python framework in the arena — ~5.5x Django, and a clear demonstration of why so many Python frameworks build on ASGI. 🐍