Read an EESSI bot FAILURE

What it does: stack bot-classify reads what an EESSI software-layer bot job left behind and says which kind of FAILURE it was: a real failure, ERROR-only noise, an empty tarball, or infra. A job that fits none of them is unclassified, for a person to read. The command never posts or edits anything, and it always exits 0.

Why you’d use it: the bot marks a job FAILURE whenever the job output holds a line matching ERROR:, FATAL:, FAILED: or required modules missing:, or lacks No missing installations or a tarball created! line. Those are the checks in bot/check-build.sh and bot/check-test.sh in EESSI/software-layer-scripts. In software-layer pull requests 1873, 1869, 1834 and 1760 every job built and packaged its software, and the FAILURE was the test step finding an ERROR: line. In 1732 and 1733 everything was already installed, so the tarball was empty. In 1741 the bot configuration kept jobs from being submitted at all. The builds that did fail were in 1818, 1823, and the first a64fx job of 1873.

The inputs are any mix of:

  • a saved bot comment, the job table the bot posts and updates;

  • the _bot_job<id>.result file the bot writes in the job directory;

  • the job output, slurm-<id>.out.

Inputs that name the same job id are read as one job. A comment that holds several job tables gives one job per table.

eb-stack stack bot-classify tests/bot_classify_fixtures/pr1873-zen2-217265.md
error-only x86_64/amd/zen2 job 217265 (eessi-bot-mc-aws): tests/bot_classify_fixtures/pr1873-zen2-217265.md
  why: `ERROR: ` matched, but No missing installations and the tarball was created
  evidence: found `No missing installations`
  evidence: eessi-2025.06-software-linux-x86_64-amd-zen2-17931241740.tar.zst: entries: 18368
  evidence: test result: found message matching ERROR:
  evidence: test result: EESSI test suite was not run, test step itself failed to execute.
real-failure=0 error-only=1 empty-tarball=0 infra=0 success=0 unclassified=0 jobs=1

How to read it:

  • The first line is the class, the CPU target, the Slurm job id, the bot instance, and the inputs.

  • why is the rule the class came from.

  • Each evidence line is a check the bot reported or a line from the job output. test result: marks the test step’s checks.

  • The count line has one field per class and the number of jobs.

The classes

Class

When

Sub-class

real-failure

FATAL:, FAILED: or required modules missing: matched; or ERROR: matched and the job never said No missing installations; or the test suite reported failing test cases

the campaign build-finding class of the job output, unknown without it

error-only

ERROR: matched in the build or test step, but the job said No missing installations and the tarball has contents

none

empty-tarball

the tarball was created with no entries, and no check found an error

none

infra

the bot submitted no job, could not merge the pull request, found no job output, or the job output shows the scheduler, the container, CernVM-FS or the network failing

bot-config, merge, missing-output, scheduler, container, network, compat-layer

success

every check passed

none

unclassified

none of the above

none

The fixtures under tests/bot_classify_fixtures are real comments, trimmed to the job table, and the first line of each names the comment it came from.

Fixture

Pull request

Class

pr1873-zen2-217265.md

1873

error-only

pr1869-zen2.md

1869

error-only

pr1834-zen2.md

1834

error-only

pr1760-riscv64-327885.md

1760

error-only

pr1741-zen5-sigma2.md

1741

error-only

pr1732-riscv64-327189.md

1732

empty-tarball

pr1733-riscv64.md

1733

empty-tarball

pr1741-zen2-206298.md

1741

empty-tarball

pr1741-no-jobs.md

1741

infra/bot-config, one job per command

pr1760-merge.md

1760

infra/merge

pr1873-a64fx.md

1873

real-failure/unknown

pr1818-riscv64-329698.md

1818

real-failure/unknown, real-failure/resource with slurm-329698.out

pr1818-riscv64-329809.md

1818

real-failure/unknown

pr1823-riscv64-first.md

1823

real-failure/unknown

pr1732-riscv64-327102.md

1732

unclassified

pr1823-riscv64-second.md

1823

unclassified

Nothing to deploy

eb-stack stack bot-classify tests/bot_classify_fixtures/pr1732-riscv64-327189.md
empty-tarball riscv64/generic job 327189 (eessi-bot-riscv-prod): tests/bot_classify_fixtures/pr1732-riscv64-327189.md
  why: No missing installations and the tarball has no entries: everything was already installed
  evidence: found `No missing installations`
  evidence: eessi-2026.06-software-linux-riscv64-generic-17908491890.tar.zst: entries: 0
real-failure=0 error-only=0 empty-tarball=1 infra=0 success=0 unclassified=0 jobs=1

The build checks passed and the test step failed, but there is nothing in the tarball to deploy. The pull request was closed for that reason.

The cause of a real failure

The bot’s comment says which checks failed, not why. The cause comes from the job output, through the same table campaign findings use (classify_build_failure). Without the job output, the sub-class is unknown:

eb-stack stack bot-classify tests/bot_classify_fixtures/pr1818-riscv64-329698.md
real-failure/unknown riscv64/generic job 329698 (eessi-bot-riscv-prod): tests/bot_classify_fixtures/pr1818-riscv64-329698.md
  why: the job output matched `FATAL:`
  evidence: found message matching FATAL:
  evidence: found message matching ERROR:
  evidence: no message matching No missing installations
real-failure=1 error-only=0 empty-tarball=0 infra=0 success=0 unclassified=0 jobs=1

Add the job output. slurm-329698.out holds the line the maintainer quoted for that job:

eb-stack stack bot-classify \
  tests/bot_classify_fixtures/pr1818-riscv64-329698.md \
  tests/bot_classify_fixtures/slurm-329698.out
real-failure/resource riscv64/generic job 329698 (eessi-bot-riscv-prod): tests/bot_classify_fixtures/pr1818-riscv64-329698.md, tests/bot_classify_fixtures/slurm-329698.out
  why: the job output matched `FATAL:`
  evidence: found message matching FATAL:
  evidence: found message matching ERROR:
  evidence: no message matching No missing installations
  evidence: OSError: [Errno 28] No space left on device
real-failure=1 error-only=0 empty-tarball=0 infra=0 success=0 unclassified=0 jobs=1

A full disk or an exceeded quota is resource, a failed EasyBuild sanity check is sanity, and a compiler error is compile. The bot’s own ERROR: and FATAL: prefixes are removed before a line is classified: they say a step failed, not which one.

Jobs that never ran

A bot instance that submits no job posts an update instead of a job table. Each command it did not submit is one infra/bot-config job:

eb-stack stack bot-classify tests/bot_classify_fixtures/pr1741-no-jobs.md
infra/bot-config x86_64/amd/zen5 (eessi-bot-sigma2): tests/bot_classify_fixtures/pr1741-no-jobs.md
  why: the bot instance submitted no job for the build command
  evidence: handling command `build architecture:x86_64/amd/zen5 repository:eessi.io-2025.06-software instance:eessi-bot-sigma2` resulted in:
  evidence: no jobs were submitted
infra/bot-config aarch64/nvidia/grace (eessi-bot-sigma2): tests/bot_classify_fixtures/pr1741-no-jobs.md
  why: the bot instance submitted no job for the build command
  evidence: handling command `build architecture:aarch64/nvidia/grace repository:eessi.io-2025.06-software instance:eessi-bot-sigma2` resulted in:
  evidence: no jobs were submitted
real-failure=0 error-only=0 empty-tarball=0 infra=2 success=0 unclassified=0 jobs=2

In the job output, these lines are infra:

  • sbatch: error:, Batch job submission failed, DUE TO NODE FAILURE: scheduler;

  • container creation failed, or a failed image pull: container;

  • Failed to initialize root file catalog, Transport endpoint is not connected, a /cvmfs/... repository that is not available!, a host name that does not resolve, or a refused connection: network;

  • a missing job configuration file, or a repository the container does not know how to mount: bot-config;

  • a Perl Can't locate ... in @INC whose @INC is under /compat/, which means the compatibility layer’s Perl ran instead of the build environment’s: compat-layer.

When the inputs do not say

eb-stack stack bot-classify tests/bot_classify_fixtures/pr1732-riscv64-327102.md
unclassified riscv64/generic job 327102 (eessi-bot-riscv-prod): tests/bot_classify_fixtures/pr1732-riscv64-327102.md
  why: the checks that failed do not say why; read the job output
  evidence: no message matching No missing installations
  evidence: No artefacts were created or found.
real-failure=0 error-only=0 empty-tarball=0 infra=0 success=0 unclassified=1 jobs=1

No check found an error, and there is no tarball. In pull request 1733 the same shape was a full disk, but the comment does not say so, and the command does not guess. Pass the job output to get a class.

JSON

--json prints the jobs and the class counts:

eb-stack stack bot-classify --json \
  tests/bot_classify_fixtures/pr1818-riscv64-329698.md \
  tests/bot_classify_fixtures/slurm-329698.out

Each job has sources, job_id, instance, repository, target, class, sub_class, reason, and evidence. counts has one key per class. The MCP tool eb_stack_bot_classify takes files and returns the text, the jobs and the counts.

Limits

  • The comment holds only the checks the bot ran. The cause of a real failure needs the job output, which stays on the bot host.

  • A job output line is classified by the campaign table, so a cause that table does not know is unknown.

  • A job that is still running has no finished row and is unclassified.