Repository navigation
Tachyon 1.4.3 removes duplicate request machinery and corrects middleware/documentation consistency while retaining the public Python API and default installation.
- One user middleware chain across HTTP, WebSocket and lifespan, with consistent registration order and Tachyon app identity.
- Documentation conflicts fail explicitly; handlers prepare before publication. Relative/mounted schema URLs remain supported.
- One body-limit authority, shared JSON response transport, unused metadata/branches removed.
- Cached
Depends(Class)on the audited 15-node graph visits 15 bindings instead of 30, retaining lifetime and override validation. This is an operation count, not a general speedup. - Version-bound public website evidence is generated from matching runtime audits; future versions cannot inherit stale benchmark labels.
Validation: 933 source/example tests per Python 3.10–3.13; 229 contracts per original and sdist-rebuilt installed wheel; 40 local release gates passed. The final load/rest trial validated 664,129 responses with zero errors in 75.3 seconds. Execution evidence is macOS ARM64; Linux/Windows and multi-hour stability are not claimed.
Compatibility corrections and the unread-upload TCP-close boundary are documented in the migration guide.
Full audit and raw measurements · Changelog
One portable wheel and one sdist are the exact locally validated pair. Publication uses the local manifest; no hosted Actions platform coverage is inferred.
Published-package comparison with FastAPI
FastAPI 0.162.0 / Pydantic 2.14.3, CPython 3.13.18, macOS ARM64, one Uvicorn asyncio/h11 worker, concurrency eight. Five repetitions; exact response validation; 240,000 measured responses including the ASGI control, zero errors.
| Scenario | Tachyon req/s | FastAPI req/s | Ratio | p99 Tachyon / FastAPI (ms) |
|---|---|---|---|---|
| hello | 13,411 | 9,242 | 1.45× | 0.703 / 0.990 |
| path_query | 11,865 | 7,411 | 1.60× | 0.762 / 1.190 |
| body | 10,240 | 7,574 | 1.35× | 1.084 / 1.187 |
| nested_body | 10,084 | 7,521 | 1.34× | 1.122 / 1.188 |
| response_model | 12,835 | 8,703 | 1.47× | 0.749 / 1.035 |
| header | 12,785 | 8,340 | 1.53× | 0.736 / 1.097 |
| dependency | 11,303 | 7,899 | 1.43× | 0.905 / 1.166 |
| query | 11,675 | 7,139 | 1.64× | 0.789 / 1.359 |
Ratios apply only to these local async scenarios; client/host contention is uncontrolled. Percentiles are medians of block percentiles, not pooled. No general multiplier or production capacity is claimed. Attached JSON retains the methodology and raw measurements.