Refuse a plan an architecture cannot build

What it does: stack arch reads an easystack as a plan. With --easyconfigs it also closes the dependencies. An entry the architecture cannot build fails the command. The message names that entry and the architecture.

Why you’d use it: a shared EESSI easystack is built on every CPU architecture. GDRCopy needs CUDA, and its header rejects riscv64. Software-layer pull request 1775 added other lines to eessi-2026.06-eb-5.4.0-2026.1.yml, and the riscv64 bot failed on the GDRCopy line that was already there. Pull request 1822 tried GDRCopy on riscv64 alone and failed the same way. Pull request 1876 failed bwa-mem2 on aarch64 because the Makefile passes -msse.

eb-stack stack arch \
  --arch riscv64 \
  --easystack tests/arch_unsupported_fixtures/easystacks/eessi-2026.06-eb-5.4.0-2026.1.yml \
  --easyconfigs tests/arch_unsupported_fixtures/easyconfigs \
  --installed tests/arch_unsupported_fixtures/installed-riscv64.txt \
  --additional-arch tests/arch_unsupported_fixtures/additional_arch/riscv64/generic/eessi-2026.06-eb-5.4.0-2026.1.yml

The command prints:

GDRCopy-2.5.2-GCCcore-15.2.0.eb is unsupported on riscv64: needs CUDA; gdrconfig.h errors with architecture is not supported (software-layer #1775, #1822)
GDRCopy-2.5.2-GCCcore-15.2.0.eb is flagged on riscv64: it has never been installed there and is not in the riscv64 additional_arch easystack
parse_failed=0 unsupported=1 flagged=1

The first line is the hard failure. GDRCopy-2.5.2-GCCcore-15.2.0.eb is the entry. riscv64 is the architecture. The reason and the pull request numbers come from data/arch-unsupported.toml. The command exits 1, and the error repeats that line.

The second line is the cheap check, and it does not read recipes. The entry is in the shared easystack, --installed has no GDRCopy/2.5.2-GCCcore-15.2.0 line, and --additional-arch does not name the file. parse_failed=, unsupported=, and flagged= are the three counts. --unsupported replaces the bundled list.

A plan that only depends on the entry names the depender. This easystack lists nvtop, and nvtop depends on GDRCopy:

eb-stack stack arch \
  --arch riscv64 \
  --easystack tests/arch_unsupported_fixtures/easystacks/nvtop-only.yml \
  --easyconfigs tests/arch_unsupported_fixtures/easyconfigs
GDRCopy-2.5.2-GCCcore-15.2.0.eb is unsupported on riscv64: needs CUDA; gdrconfig.h errors with architecture is not supported (software-layer #1775, #1822) (required by nvtop-3.3.2-GCCcore-15.2.0.eb)
parse_failed=0 unsupported=1 flagged=0

(required by nvtop-3.3.2-GCCcore-15.2.0.eb) means the plan did not name GDRCopy. The closure did. flagged=0 is the cheap check staying quiet, because --installed was not passed.

bwa-mem2 is the aarch64 entry:

eb-stack stack arch \
  --arch aarch64 \
  --easystack tests/arch_unsupported_fixtures/easystacks/bwa-mem2.yml \
  --easyconfigs tests/arch_unsupported_fixtures/easyconfigs
bwa-mem2-2.3-GCC-14.3.0.eb is unsupported on aarch64: the Makefile passes -msse and -msse2, which aarch64 gcc rejects (software-layer #1876)
parse_failed=0 unsupported=1 flagged=0

--arch aarch64/neoverse_v1 is the same family as --arch aarch64. The family is the text before the first slash. With no --arch, every family in the list is checked. --arch x86_64 does not reject GDRCopy: the list has no x8664entry for it.

The command also parses each easyconfig for each target architecture (x86_64, aarch64, riscv64 by default, or the families named with --arch). A file named in the plan that parses on one architecture but not another is reported before any bot build. Software-layer pull request 1798 failed this way: Go-1.26.7.eb indexes sources by ARCH and has no riscv64 entry. The parser does not evaluate that lookup, so the report is that sources could not be read:

eb-stack stack arch \
  --arch riscv64 \
  --easystack tests/arch_unsupported_fixtures/easystacks/go-1.26.7.yml \
  --easyconfigs tests/arch_unsupported_fixtures/easyconfigs
Go-1.26.7.eb failed to parse on riscv64: parse tests/arch_unsupported_fixtures/easyconfigs/g/Go/Go-1.26.7.eb: sources is assigned but could not be read, so planning with this recipe would get its identity, dependencies, or sources wrong
parse_failed=1 unsupported=0 flagged=0

An entry that is in the additional_arch easystack is not flagged, even when it has never been installed. It is still unsupported when the list says so. Passing --installed with more than one --arch needs a prefix, riscv64=installed-riscv64.txt. The same prefix form works for --additional-arch. A directory is every easystack under it.

stack batch uses the same list before it writes a plan. A root that includes GDRCopy, or depends on it, exits 1 and does not write plan.json. --arch limits that check to one family. With no --arch, every family in the list is checked, because the batch is a shared easystack.

Software-layer pull request 1623 added waLBerla 7.2. The bot failed on aarch64 neoversev1, a64fx, and nvidia/grace because SIMDAlignment() was not constexpr under SVE. The from-commit pin a69e693 patches that, and the bot then succeeded on those targets. waLBerla is not in the list. The cheap check is what flags a shared entry before that first bot run, while it has never been installed on the architecture and is not in that architecture’s additional_arch easystack.

Per-architecture source probe

What it does: --probe-sources resolves each planned easyconfig’s source_urls and sources on every architecture family the command checks, then probes every concrete URL with HEAD and a ranged GET fallback (Range: bytes=0-0). The families are the ones --arch names, with a target such as aarch64/neoverse_v1 counted as aarch64. With no --arch they are x86_64, aarch64, and riscv64. A file already reported as parse_failed on a family is not probed there, and a family --arch does not name is neither parsed nor probed.

Why you’d use it: software-layer pull request 1670 failed because the Kaleido wheel naming differed by architecture. The plan parsed and closed, then the bot burned a full round failing to fetch source files on non-x86 targets. The probe catches that before the bot run.

eb-stack stack arch \
  --easystack tests/source_probe_fixtures/easystacks/kaleido.yml \
  --easyconfigs tests/source_probe_fixtures/easyconfigs \
  --probe-sources

The fixture’s source_urls is __BASE_URL__/. tests/source_probe.rs and tests/doc_walkthrough.rs copy the tree, put the address of a local server that has only the x8664wheel in place of __BASE_URL__, and run this command on the copy. With the address written back as __BASE_URL__, the command prints:

Kaleido-0.2.1-GCCcore-15.2.0.eb source kaleido-0.2.1-manylinux2014-aarch64.whl is arch-missing on aarch64: __BASE_URL__/kaleido-0.2.1-manylinux2014-aarch64.whl (attempt 1/3 HEAD=status:404 RANGE=status:404)
Kaleido-0.2.1-GCCcore-15.2.0.eb source kaleido-0.2.1-manylinux2014-riscv64.whl is arch-missing on riscv64: __BASE_URL__/kaleido-0.2.1-manylinux2014-riscv64.whl (attempt 1/3 HEAD=status:404 RANGE=status:404)
parse_failed=0 unsupported=0 flagged=0 source_unreachable=0 source_arch_missing=2

How to read it: each line names the easyconfig, the concrete source file for that architecture, the URL that was tried, and the HTTP status of the HEAD and the ranged GET. arch-missing means another probed family fetched this source slot and this family did not, so the recipe names a file that exists for one architecture and not this one. unreachable means no probed family fetched it. source_unreachable and source_arch_missing in the count line count the two classes. Any finding exits 1. --probe-retries (default 2) and --probe-timeout-s (default 8) tune the network check.

One architecture

What it does: --arch aarch64 --probe-sources probes the plan on aarch64 only.

Why you’d use it: a pull request that adds an architecture-specific easyconfig, or one bot target that failed to fetch, needs only that family checked. A recipe that does not parse on another family does not stop the probe.

eb-stack stack arch \
  --arch aarch64 \
  --easystack tests/source_probe_fixtures/easystacks/kaleido.yml \
  --easyconfigs tests/source_probe_fixtures/easyconfigs \
  --probe-sources

On the same copy and server it prints:

Kaleido-0.2.1-GCCcore-15.2.0.eb source kaleido-0.2.1-manylinux2014-aarch64.whl is unreachable on aarch64: __BASE_URL__/kaleido-0.2.1-manylinux2014-aarch64.whl (attempt 1/3 HEAD=status:404 RANGE=status:404)
parse_failed=0 unsupported=0 flagged=0 source_unreachable=1 source_arch_missing=0

How to read it: with one family probed there is no other architecture to compare with, so a source that family cannot fetch is unreachable, not arch-missing. The line still names the file, the URL, and the 404, and the command exits 1. Run without --arch to learn whether the same source slot resolves to a file that exists on another architecture. --arch x86_64 on this fixture probes only the x8664wheel, which the server has, and exits 0.

Two architectures

What it does: --arch x86_64 --arch aarch64 --probe-sources probes those two families and no other. A target such as x86_64/amd/zen4 is still the family before the slash. Naming a family and one of its targets, --arch aarch64 --arch aarch64/neoverse_v1, probes that family once. The order of the flags does not change the report.

Why you’d use it: two bot targets failed the fetch and a third was not in the run. The comparison is between the families that were named. A family left off the command is not fetched, and it does not turn arch-missing into unreachable or the other way around.

eb-stack stack arch \
  --arch x86_64 \
  --arch aarch64 \
  --easystack tests/source_probe_fixtures/easystacks/kaleido.yml \
  --easyconfigs tests/source_probe_fixtures/easyconfigs \
  --probe-sources

On the same copy and server it prints:

Kaleido-0.2.1-GCCcore-15.2.0.eb source kaleido-0.2.1-manylinux2014-aarch64.whl is arch-missing on aarch64: __BASE_URL__/kaleido-0.2.1-manylinux2014-aarch64.whl (attempt 1/3 HEAD=status:404 RANGE=status:404)
parse_failed=0 unsupported=0 flagged=0 source_unreachable=0 source_arch_missing=1

How to read it: the x8664wheel is on the server, so the aarch64 wheel is arch-missing. riscv64 is not named and is not probed. The same command with --arch aarch64 --arch riscv64 prints both wheels as unreachable, with source_unreachable=2 and source_arch_missing=0, because neither probed family fetched the source slot. The x8664wheel is not consulted.

See the command-line reference.