Files
DECNET/tests/perf/README.md
anti 319c1dbb61 added: profiling toolchain (py-spy, pyinstrument, pytest-benchmark, memray, snakeviz)
New `profile` optional-deps group, opt-in Pyinstrument ASGI middleware
gated by DECNET_PROFILE_REQUESTS, bench marker + tests/perf/ micro-benchmarks
for repository hot paths, and scripts/profile/ helpers for py-spy/cProfile/memray.
2026-04-17 13:13:00 -04:00

2.0 KiB

DECNET Profiling

Five complementary lenses. Pick whichever answers the question you have.

1. Whole-process sampling — py-spy

Attach to a running API and record a flamegraph for 30s. Requires sudo (Linux ptrace scope).

./scripts/profile/pyspy-attach.sh               # auto-finds uvicorn pid
sudo py-spy record -o profile.svg -p <PID> -d 30 --subprocesses

If py-spy "doesn't work", it is almost always one of:

  • Attached to the Typer CLI PID, not the uvicorn worker PID (use pgrep -f 'uvicorn decnet.web.api').
  • kernel.yama.ptrace_scope=1 — run with sudo or sudo sysctl kernel.yama.ptrace_scope=0.
  • The API isn't actually running (a --dry-run deploy starts nothing).

2. Per-request flamegraphs — Pyinstrument

Set the env flag, hit endpoints, find HTML flamegraphs under ./profiles/.

DECNET_PROFILE_REQUESTS=true decnet deploy --mode unihost --deckies 1
# in another shell:
curl http://127.0.0.1:8000/api/v1/health
open profiles/*.html

Off by default — zero overhead when the flag is unset.

3. Deterministic call graph — cProfile + snakeviz

For one-shot profiling of CLI commands or scripts.

./scripts/profile/cprofile-cli.sh services      # profiles `decnet services`
snakeviz profiles/cprofile.prof

4. Micro-benchmarks — pytest-benchmark

Regression-gate repository hot paths.

pytest -m bench tests/perf/ -n0                 # SQLite backend (default)
DECNET_DB_TYPE=mysql pytest -m bench tests/perf/ -n0

Note: -n0 disables xdist. pytest-benchmark refuses to measure under parallel workers, which is the project default (-n logical --dist loadscope).

5. Memory allocation — memray

Hunt leaks and allocation hot spots in the API / workers.

./scripts/profile/memray-api.sh                 # runs uvicorn under memray
memray flamegraph profiles/memray.bin

Load generation

Pair any of the in-process lenses (2, 5) with Locust for realistic traffic:

pytest -m stress tests/stress/