datadog/debug-build-times
> Diagnose and fix slow base venv build times caused by unnecessary recompilation of native extensions (CMake, Cython, Rust) across riot generate runs. Use when CI base venv builds are slow, when ext_cache isn't saving time, or when investigating warm build regressions.
npx skills add https://github.com/DataDog/dd-trace-py --skill debug-build-times
For architecture, root causes, key files, and env vars see
docs/build_system.rst — "Debugging Build Performance" section.
On a warm run (extensions cached) pip install -e . should complete in under 30s.
Create a throwaway test_venv_cache.sh in the repo root (don't commit it), run it,
then inspect the metadata files:
bash test_venv_cache.sh 3.13 # or whichever Python version
cat debug_ext_metadata_cold.txt
cat debug_ext_metadata_warm.txt
#!/usr/bin/env bash
set -euo pipefail
PYTHON_VERSION="${1:-3.12}"
CACHE_ROOT="/tmp/dd_ext_cache_test"
# IMPORTANT: use a version-specific venv to match CI's ext_cache_venv${PYTHON_VERSION} pattern.
# ext_cache.py uses sys.executable to call setup.py ext_hashes, which computes extension
# targets with the running Python's suffix (e.g. cpython-313). If the venv Python doesn't
# match the build Python, cache save/restore silently operates on the wrong suffix and the
# ext_cache becomes a no-op for that Python version.
EXT_CACHE_VENV="/tmp/dd_ext_cache_venv${PYTHON_VERSION}"
METADATA_COLD="debug_ext_metadata_cold.txt"
METADATA_WARM="debug_ext_metadata_warm.txt"
header() { echo; echo "══════════════════════════════════════════"; echo " $*"; echo "══════════════════════════════════════════"; }
step() { echo "── $*"; }
# One-time venv for running ext_cache.py (mirrors CI's EXT_CACHE_VENV)
if [ ! -d "$EXT_CACHE_VENV" ]; then
step "Creating ext_cache venv (python${PYTHON_VERSION})"
"python${PYTHON_VERSION}" -m venv "$EXT_CACHE_VENV"
"$EXT_CACHE_VENV/bin/pip" install -q wheel cmake setuptools_rust Cython
fi
cache_restore() { "$EXT_CACHE_VENV/bin/python" scripts/ext_cache.py --root "$CACHE_ROOT" restore; }
cache_save() { "$EXT_CACHE_VENV/bin/python" scripts/ext_cache.py --root "$CACHE_ROOT" save; }
run_build() {
local label="$1" metadata_out="$2"
header "$label build (python $PYTHON_VERSION)"
step "Restoring from cache"; cache_restore
step "riot generate"
export _DD_DEBUG_EXT=1 _DD_DEBUG_EXT_FILE="$metadata_out"
riot -P generate --python="$PYTHON_VERSION"
unset _DD_DEBUG_EXT _DD_DEBUG_EXT_FILE
step "Saving to cache"; cache_save
cat "$metadata_out" 2>/dev/null || echo "(no metadata written)"
}
header "Clearing test state"
rm -rf ".riot/venv_py${PYTHON_VERSION/./}"*
rm -rf "$CACHE_ROOT"
rm -rf .eggs # stale eggs cause ENOTEMPTY on Python ≤3.10's legacy setuptools
# Remove compiled .so files for a true cold state
find ddtrace -name "*.so" -o -name "*.dylib" -o -name "*.pyd" | grep -v "_vendor" | xargs rm -f 2>/dev/null || true
run_build "COLD" "$METADATA_COLD"
header "Clearing riot venv for warm run"
rm -rf ".riot/venv_py${PYTHON_VERSION/./}"*
run_build "WARM" "$METADATA_WARM"
header "Summary"
diff --unified=0 --label COLD --label WARM "$METADATA_COLD" "$METADATA_WARM" || true
Expected results: Cold ~100s, warm <10s. Any extension showing >0.05s on warm is
rebuilding unnecessarily.
Run bash test_venv_cache.sh 3.13 with _DD_DEBUG_EXT=1 (the script sets this). The
metadata files show per-phase and per-extension times.
In debug_ext_metadata_warm.txt, any extension with >0.05s is rebuilding.
print(f"DEBUG {ext.name}: ext_path={ext_path} exists={ext_path.exists()} needs_rebuild={needs_rebuild}")
.riot/venv_py3131/bin/pip --disable-pip-version-check install -e . -v 2>&1 | grep "DEBUG\|skipping\|building"
# In a Python shell after restoring from cache:
from pathlib import Path
from distutils.dep_util import newer_group
import sysconfig
suffix = sysconfig.get_config_var('EXT_SUFFIX')
source_dir = Path('ddtrace/profiling/collector') # adjust as needed
ext_path = source_dir / f'_memalloc{suffix}'
srcs = [s for s in source_dir.glob('**/*') if s.is_file() and s.suffix not in {'.py','.pyc','.pyi','.c','.so','.dylib','.dll','.pyd'}]
ext_mtime = ext_path.stat().st_mtime
newer = [(str(s), round(s.stat().st_mtime - ext_mtime, 3)) for s in srcs if s.stat().st_mtime > ext_mtime]
print(newer)
/tmp/dd_ext_cache_venv3.13/bin/python scripts/ext_cache.py --root /tmp/dd_ext_cache_test restore 2>&1 | grep -i "warning\|error\|restoring"
Warnings about "No cached files found" mean the hash changed between save and restore.
Run python setup.py ext_hashes --inplace twice and compare — if hashes differ, there's
a hash stability bug.
Take datadog/debug-build-times from the repository into ~/.claude/skills for personal
use, or into .claude/skills inside a project.
The agent identifies a skill by the name field in its header. Two skills with the
same name cannot sit side by side — one of them will be ignored.
The instructions reference pip.
Without those the skill loads but fails at the first command.