Start with the application, not the integration backlog

A serious service needs validation, identity, PostgreSQL, policy, background work, observability and somewhere to run. In Wreath these are declarations over one application lifecycle, not middleware assembled around several competing request models.

This route gets compiled binding, bounded fields, native JSON, replay-safe writes, request IDs, security headers and an OpenAPI operation from the same declarations:

from dataclasses import dataclass
from typing import Annotated

from wreath import Request, Wreath
from wreath.binding import Body, Field
from wreath.policy import HttpPolicy, IdempotencyPolicy
from wreath.policy.request_id import RequestIdPolicy
from wreath.policy.security import SecurityHeadersPolicy


@dataclass
class RunRequest:
    prompt: Annotated[str, Field(min_length=3, max_length=2_000)]
    model: Annotated[str, Field(min_length=2, max_length=80)]


app = Wreath(
    http_policy=HttpPolicy(
        idempotency=IdempotencyPolicy(),
        request_id=RequestIdPolicy(),
        security_headers=SecurityHeadersPolicy(),
    )
)


@app.post("/runs")
async def create_run(
    request: Request,
    command: Annotated[RunRequest, Body()],
) -> dict:
    return {"state": "queued", "model": command.model, "prompt": command.prompt}


app.enable_api_docs(environments=("development",), try_it_out=True)

uv add wreath
uv run wreath dev app:app

The framework stays ordinary ASGI, so uvicorn app:app remains valid. The native server is available when you want Wreath's complete HTTP, telemetry and recording path.

Start where your system gets hard

The documentation builds nine complete systems around the pressure that usually forces a second stack:

Underneath those stories are the conventional guides: HTTP and OpenAPI, data and migrations, CLI tasks, and the production runbook. The camera-trap service is the canonical runnable application. Wreath is pre-1.0; the version, platform and upgrade contract states exactly what the current docs cover.

Performance claims come with the workload

Wreath moves repeated work to startup and byte-heavy work into owned native kernels. The comparisons below use retired userspace instructions, equivalent verified output, alternating samples and retained raw evidence. They do not turn one elapsed-time run into a throughput claim.

Native gzip, shaped to the document

Wreath owns both halves of gzip. The encoder and decoder are independent native kernels: they share the RFC 1951/1952 wire format, not one duplex algorithm. When Content-Type is known, JSON, GraphQL, HTML, logs, plain text, and high-entropy JSON take separately compiled policies. The result is still ordinary gzip readable by browsers, zlib, zlib-ng, and libdeflate.

Retired userspace instructions per uncompressed byte at level 6, lower is better. Each cell is Wreath / libdeflate / zlib-ng:

That is fewer encode and decode instructions in every document row. The cache figures are the 500 kB representatives: the high-entropy encoder retires only 4.13 LLC misses/kB, while the regular-document decoder stays at 1.79–2.22 LLC misses/kB with a 97.86–98.17% LLC hit rate. Across the corpus, Wreath's output is 0.33% larger than zlib-ng's and 0.80% larger than libdeflate's—the ratio cost kept beside the instruction win rather than hidden.

The comparison uses repository-native boundaries, application- or stream-owned warm contexts, six document sizes from 10 kB through 1 MB (the high-entropy case is 500 kB), and hardware performance counters on Linux x86-64 / Ryzen 7 7730U. A subsequent cache sweep over 30 JSON, GraphQL, HTML, text, and log documents chose the 32 Ki-entry format hash table and shallower per-format search limits: encode instructions fell 4.79%, L1D misses 12.87%, and LLC misses 52.07% on the same corpus. The shipped balance retires 49.04 encode instructions/byte with 91.62% L1D and 93.92% LLC hit rates; output is 23.71% of input. No elapsed time, cycles, or IPC enters either result. The repository retains its measurement data with the benchmark sources.

One real request, all the way through

This is not a plaintext race or a synthetic JSON payload. All three configurations render the same operations-intelligence dashboard from one successful request: HTTP policy and sessions, nested typed input, bearer authentication, Cedar, overlapping PostgreSQL and HTTP wire calls, eleven requested rows projected from a sparse 730-day × 48-tenant × 6-measure source, eleven chart paths, temporal, geospatial and vector work, ranked pagination, protobuf and MessagePack exports, an escaped template, compression, and HTML emission.

Retired userspace instructions per request, lower is better. Five samples are too few for a meaningful histogram, and scaling every row independently made the old glyphs impossible to compare. This instead prints the observed range, its width relative to the median, and the median directly:

The cache result is clearest in absolute terms: optimized Wreath records 159,258 L2 misses per request. The other stacks record 2.24–2.81 million— 14.06× to 17.65× more misses per request.

The normalized Wreath ratio looks high because Wreath retires roughly 52× fewer instructions while some cache traffic remains fixed. Dividing that traffic by the much smaller instruction count raises misses per thousand instructions; it does not mean more cache misses per request. The retained samples include the full L1 and AMD L2 demand and prefetch counters.

Format-aware Wreath retired 28.96× fewer instructions than FastAPI in this run. The dictionary-aware configuration retired 52.08× fewer instructions, removing 3,222,669 instructions per request—a further 44.38% from Wreath's full request path. BlackSheep on Granian was the least instruction-heavy of the three ecosystem stacks, at 28.85× and 51.88× the two Wreath medians. The DCZ response is RFC 9842 dcz: a real response for neighbouring resource 41 is the client-held dictionary for resource 42, whose 35,805-byte body differs at only two bytes. The same configuration falls back to standard concatenated gzip members for ordinary clients, recompressing the 249-byte dynamic prefix while reusing the independently readable stable member. It does not claim that DCZ and gzip are nested in one response.

All five rows use TLS 1.3 over HTTP/1.1 and are medians of five alternating 30/15-request slopes. The unchanged A/A controls differed from the primary medians by 0.58%, 0.90%, 0.14%, 0.44%, and 0.11%, respectively. The harness accepts a DCZ sample only after checking its dictionary hash, decoded response facts, secure transport, and both required Vary fields; it separately exercises and decodes the fragment-gzip fallback before attaching counters.

The ecosystem implementations are not slow pure-Python foils. FastAPI uses Starlette policy middleware, Pydantic, and Uvicorn/uvloop/httptools. Sanic uses its native production server in single-process mode. BlackSheep uses Granian's single-threaded ASGI runtime with uvloop, the faster small-process mode its server documents. Sanic and BlackSheep share one msgspec-typed business kernel; all three use cedarpy, asyncpg, aiohttp, NumPy, Jinja, protobuf, msgspec, and standard gzip. Wreath performs the corresponding transaction through its dependency-free declarative surfaces and native data kernels.

To expose the successful-path application cost layer by layer, a smaller companion request removes the dashboard calculations while keeping the same routing, CORS, validation, auth, Cedar, PostgreSQL, HTTP, and JSON result. The starred cells below are deliberately called out because they are not like-for-like framework feature costs:

The cumulative decomposition shows where work enters:

^* Sanic and BlackSheep share a hand-written, success-path adapter: msgspec decodes the known-good body, the query is converted inline, and bearer "authentication" is an exact comparison with "Bearer user". Wreath's cells exercise its public binder and bearer backend, including structured validation refusals, case-insensitive scheme parsing, duplicate-credential refusal, Identity publication, protected-route resolution, and a proper 401 Bearer challenge. Invalid input is therefore not equivalent: for example, the Sanic adapter answers 500 for a bad query or missing token where Wreath answers 422 or 401, and it accepts a duplicated Authorization field when the first value is valid. Every later cumulative BlackSheep and Sanic cell inherits this shortcut. Those full rows remain an account of the exact successful applications, but the starred increments are not evidence that equivalent framework validation or authentication is cheaper. In the retained Sanic run, the 1,966-instruction authentication increment is also below its available 2,747-instruction A/A resolution and is unresolved.

On this stripped companion, Wreath retired 8.43× fewer instructions than FastAPI and 5.78× fewer than both Sanic and BlackSheep. BlackSheep has the lowest ecosystem route floor; the two lightweight stacks converge once the same Cedar, PostgreSQL and outbound HTTP work dominates. Its purpose is decomposition; the holistic request above is the headline system comparison.

CORS is Starlette's CORSMiddleware, which FastAPI already brings in. It is not padded into the stack as another package. cedarpy 4.8.7 has a stateless public authorization call and no reusable compiled-policy handle. The measured request therefore includes that public lifecycle.

Both clients speak their real wire protocols to the same deterministic, in-process PostgreSQL and HTTP peers. The verifier rejects a sample unless the security, CORS, session, compression, and business response facts agree. Imports, startup, compilation, pool creation, and warm-up cancel through N/N/2 slopes. Server and generator are pinned separately with PYTHONHASHSEED=0. Instructions and the four L1 events share one non-multiplexed perf pass; five AMD component events split demand from prefetch L2 traffic in a second pass. L1 hits are accesses minus misses; aggregate L2 misses are retained only as an absolute and per-instruction normalization, not presented as one hit rate. The result contains no elapsed time, cycles, or IPC. It was recorded on CPython 3.14.7 and Linux x86-64, on a Ryzen 7 7730U. SIMD dispatch can make another architecture retire a different count.

Reproduce it:

uv sync --inexact --group benchmark
uv run python -m benchmarks.bench_holistic_stack_instructions \
  --requests 30 --trials 5 --connections 8 --warmup 16 \
  --output benchmarks/baselines/e2e-holistic-stack-instructions.json

# The cumulative framework-layer control
uv run python -m benchmarks.bench_e2e_instructions \
  --requests 4000 --trials 5 --connections 32 --warmup 500 \
  --output benchmarks/baselines/e2e-stack-instructions.json

The repository retains the holistic raw samples, Wreath application, FastAPI application, and counter harness, plus the cumulative control samples.

Install

pip install wreath
# or
uv add wreath

uv add 'wreath[linux]'     # io_uring reactor and native TLS on Linux
uv add 'wreath[h3]'        # HTTP/3; `http3` is an alias

The base wheel includes Wreath's portable C implementation. The framework remains a conforming ASGI application when you use another server.

The engineering rule

Wreath moves repeated work out of requests and moves byte-heavy work into native kernels. It does not call an idea faster because it was rewritten in C. Every performance change starts with a measurement, keeps its controls, and reports the machine and method that produced the result.

The same refusal applies to correctness. Unsupported declarations fail at startup. Hot-path complexity has executable probes. The test runner samples declared controls and asks whether the suite notices when one disappears.

Work on Wreath

uv sync
uv run wreath-check
uv run wreath test

The repository's AGENTS.md is the engineering contract.