Runs the table pages on 3.15.0b4t with parallel and sequential extras,
1-8 SQL threads, and 1 or 8 requests in flight. On the GIL build parallel
extras are slower because the SQL worker threads contend for the GIL on
every fetched row. On the free-threaded build that penalty disappears and
throughput under load is 2.2-3.1x higher. Adds a patch that resolves
extras in parallel only when the GIL is disabled, the stress and gather
microbenchmark tools, and the raw results.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F8mEgRS8swZfYqwqrVMeo5
Real datasette serve runs (uvicorn, keep-alive load, 5 interleaved rounds)
comparing cf113a5 with the combined patches, with and without httptools
and the access log. Adds the load generator and server harness.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F8mEgRS8swZfYqwqrVMeo5
Tools for profiling Datasette's request/response cycle in-process with
Python 3.15's profiling.sampling, plus the patches proposed by the study
(one per optimisation, and ALL-combined.diff). The patches are not
applied to the code. A report with measurements follows.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F8mEgRS8swZfYqwqrVMeo5
