Site Python extras on EESSI¶
Turn a PyPI package or requirements.txt into an easyconfig that installs
on top of an EasyBuild robot. Inspect records easyblock = PythonBundle,
because ingest does not yet know whether the package brings extensions.
A plan of one package emits PythonPackage with its own sources and does
not write exts_list. Requirements the robot already provides become
dependencies on the providing bundle (soupsieve via
Python-bundle-PyPI, numpy via SciPy-bundle). A plan that still has
leftovers emits a PythonBundle and keeps only those leftovers in
exts_list.
Do not pip install --target and prepend PYTHONPATH. That shadows
toolchain-built extensions and has no claim on the European Environment for Scientific Software Installations (EESSI) compat layer.
Inputs¶
Warehouse-shaped JSON (
info+urls) or arequirements.txt;the EESSI robot tree (or a fixture that mirrors its provides), passed as
--easyconfigs. It is the easyconfigs directory bundled with EESSI’s EasyBuild:$EBROOTEASYBUILD/easybuild/easyconfigsafter loading EESSI’sEasyBuildmodule;a stack policy that locks
Python,SciPy-bundle, andPython-bundle-PyPIto the versions in the robot. Seeexamples/stacks/eessi-python-extras.toml.
Inspect does not need a robot. Plan does.
Inspect¶
eb-stack package inspect \
--source fixtures/foreign_ingest/pypi_bs4/pypi.json \
--format pypi \
--toolchain GCCcore/15.2.0 \
--out-dir /tmp/bs4-inspect
package.plan.json has origin = pypi and easyblock = PythonBundle.
Optional extras (extra ==) become residuals and are not solved.
A requirements.txt is also accepted. The first spec is the root
package; later specs are additional run dependencies.
Plan¶
Source env.sh from scripts/fetch-upstream-trees.sh first.
EB_EASYCONFIGS is the develop easyconfigs tree.
Pick the toolchain the way EESSI does. A pure-Python package goes on
GCCcore, where EESSI 2026.06 installs Python, Python-bundle-PyPI and
BeautifulSoup, so the bs4 plan uses GCCcore/15.2.0 and
examples/stacks/eessi-python-gcccore.toml. A package that needs numpy
needs SciPy-bundle, which EESSI installs on gfbf/2026.1; the numpy
example below uses foss/2026.1, which includes it, and
examples/stacks/eessi-python-extras.toml. The policy’s toolchain must
be the plan’s --toolchain.
eb-stack package plan \
--source fixtures/foreign_ingest/pypi_bs4/pypi.json \
--format pypi \
--toolchain GCCcore/15.2.0 \
--easyconfigs fixtures/foreign_ingest/pypi_bs4/robot \
--stack-policy examples/stacks/eessi-python-gcccore.toml \
--out-dir /tmp/eb-stack/bundles/bs4
The lock collapses soupsieve to Python-bundle-PyPI. The emitted
recipe is PythonPackage, depends on Python 3.14.2 and
Python-bundle-PyPI 2026.04, and has no exts_list. The checksum is
the one in the fixture. The plan’s easyblock stays PythonBundle;
that class is what a multi-package overlay emits. A PythonBundle
recipe is for a set installed together, not for a single package.
Against easybuild-easyconfigs develop 981d2123dd, with --easyconfigs "$EB_EASYCONFIGS"
after sourcing env.sh, the same package is already
provided and the command does not emit a recipe. The manifest residual
is:
beautifulsoup4 4.12.3 is already provided by BeautifulSoup 4.14.3 (beautifulsoup4 4.14.3); do not emit a pip overlay
The evidence path is b/BeautifulSoup/BeautifulSoup-4.14.3-GCCcore-15.2.0.eb.
The extra-only lxml requirement stays a parse residual.
eb-stack package plan \
--source fixtures/foreign_ingest/pypi_numpy/numpy.json \
--format pypi \
--toolchain foss/2026.1 \
--easyconfigs fixtures/foreign_ingest/pypi_numpy/robot \
--stack-policy examples/stacks/eessi-python-extras.toml \
--out-dir /tmp/eb-stack/bundles/numpy
Asking for numpy against a robot that already provides it writes no
recipe. package.plan.json has an already-provided residual:
numpy 2.5.2 is already provided by SciPy-bundle 2025.06 (numpy 2.3.1).
A recipe that depends on numpy lists SciPy-bundle, not a pip numpy.
If the robot has no SciPy-bundle / PyTorch, the plan refuses rather
than emitting a wheel overlay.
A leftover that is not in EasyBuild yet (eon-akmc) keeps unsatisfied
PyPI run dependencies in the same exts_list. numpy still maps to
SciPy-bundle. A maturin/PyO3 leftover such as readcon is closed from
a Cargo source root (--cargo-source / kind = "cargo") into a
companion PythonPackage (implicit Rust / maturin / binutils)
instead of a pip extension. Pass the crate metadata next to the PyPI
plan; do not pip install the crate into the overlay.
Resolvo picks every robot-provided name. Extra site packages that Warehouse
omits (measured PyYAML for import eon) go in the ingest as a second
Warehouse object in a JSON array, or as a later requirements.txt line.
A registry name (--source eon-akmc) fetches Warehouse once and writes
ingest/pypi/<name>-<version>.json. Replan from that dump; the Boolean satisfiability (SAT) solve does not
hit the network. Meson wrap natives come from subprojects/*.wrap in
the sdist tree next to the dump (or from build_system.requires). Do not
edit the emitted .eb to add those names: put the modules in
--easyconfigs and let the lock select them. An undeclared import is
a residual, not a silent SAT edge. A PEP 518 build_system
block with build-backend = mesonpy adds the listed build requirements
and emits wrap_mode=default plus the same cargo-on-EESSI isolation
prelude used by Cargo leftovers: drop host rustc wrappers, set
LINKER="$CC" (an empty CC becomes gcc, and the value is not
evaluated), put the EESSI compat ld on PATH, and point
RUSTFLAGS -B at that ld. The cargo target triple and compat arch
come from the build host (rustc -vV, uname -m), not from a
plan-time x8664literal. Hatchling and meson-python already on
Python-bundle-PyPI stay modules; they are not pip extras.
Dependencies that state no version¶
A Python project routinely requires another package with no version. An
exts_list entry still needs one, so planning stops and asks for an index.
On this side the index is a pinned requirements file, matching what
pip freeze writes:
pip freeze > constraints.txt
eb-stack package plan ... --package-index constraints.txt
One name==version per line; comments and blank lines are ignored, and a line
that names a range rather than a version supplies nothing, since a range is not
a version. The same flag takes a Comprehensive R Archive Network (CRAN) PACKAGES file for the R side; the
format is detected from the content.
Install with the local target¶
Plan on the host. Install through campaign run:
scripts/fetch-upstream-trees.sh
set -a
source "${EB_UPSTREAM_CACHE:-$HOME/.cache/eb-stack/upstream}/env.sh"
set +a
mkdir -p /tmp/eb-stack/robot /tmp/eb-stack/bundles
rsync -a "$EB_EASYCONFIGS"/ \
/tmp/eb-stack/robot/
podman build -t eb-stack-rocky9 \
-f skills/new-package/container/rocky9/Containerfile \
skills/new-package/container/rocky9
eb-stack campaign run \
--bundle /tmp/eb-stack/bundles/bs4 \
--config examples/targets/local-podman.toml \
--target local-rocky9 \
--state /tmp/eb-stack/bundles/bs4.campaign.json
examples/targets/eessi-extend.toml is the target whose runtime command
is eessi_container.sh. The local target above is the one this checkout
runs. That EESSI target execs eb-local with EB_LOCAL_EESSI_EXTEND=1,
which loads EESSI-extend and execs eb.
User one-offs stay as python -m venv --system-site-packages on EESSI
Python. This workflow is for a shared site or project extra.