Glossary

Core hpc-compose terms, in one place. The short version of this list also appears on the Overview page; this page is the fuller reference.

One-line definitions; follow the link for the owning reference section.

allocation: The single Slurm job where all of an application's services run; one spec compiles to one allocation. See Execution Model.
artifact bundle: A named group of output paths declared under x-slurm.artifacts and exported with hpc-compose artifacts. See x-slurm.artifacts.
canary: A short, minimized probe run from hpc-compose germinate that writes latest-canary.json and leaves latest.json untouched. See germinate.
cache directory: Shared storage for imported and prepared images, visible from both the submission host and the compute nodes. See x-slurm.cache_dir.
compose file / spec: The YAML file describing services, runtime backend, and Slurm settings; "spec" and "compose file" are the same thing. See Spec Reference.
context: The resolved view of settings, profile, binaries, interpolation variables, and runtime paths for an invocation. See context.
failure policy: Per-service restart behavior under services.<name>.x-slurm.failure_policy. See failure_policy.
local mode: Running a plan on the current Linux host through the local Pyxis/Enroot supervisor instead of submitting to Slurm; single-host and Pyxis-only. See up --local.
login node / submission host: The host where you run hpc-compose and from which jobs are submitted; "login node" and "submission host" name the same machine. See Operate a Real Cluster Run.
preflight: Checks of local tools, paths, backend support, and optional cluster profiles before a run. See preflight.
prepare: The login-node phase that imports base images and builds prepared runtime artifacts, reused later by up and run. See x-runtime.prepare.
profile: A named settings block in .hpc-compose/settings.toml, selected with --profile <name>. See Common Flags.
readiness: A gate that holds a dependent service until a probe passes; types are sleep, tcp, http, and log. See readiness.
rendezvous: Same-cluster service discovery through JSON records under the shared cache directory; not DNS, auth, or a service mesh. See x-slurm.rendezvous.
resume: Resume-aware reruns backed by a shared x-slurm.resume.path and attempt-aware state. See x-slurm.resume.
right-sizing: Comparing requested versus observed usage to suggest reductions (inspect --rightsize) plus the efficiency grade from score. See Tracked Runtime.
runtime backend: The mechanism used to launch services: Pyxis/Enroot, Apptainer, Singularity, or host software, selected with runtime.backend. See runtime.
service: One container or host process in the allocation, defined under services.<name> (steps is an accepted alias). See Service fields.
smoke test: A finite end-to-end run (hpc-compose test) where every service must start, pass readiness, and complete successfully. See test.
sweep: An embedded sweep block expanded by hpc-compose sweep submit into many independent tracked allocations, one per trial. See sweep.
tracked job / tracked run: Metadata under .hpc-compose/<job-id>/ that lets status, ps, watch, logs, stats, and artifacts reconnect to a run later; "tracked job" and "tracked run" are the same thing. See Tracked Runtime.
x-runtime.prepare: The spec block for image-preparation commands and mounts; x-enroot.prepare is an accepted Pyxis/Enroot alias. See x-runtime.prepare.
x-slurm: The spec section for Slurm settings and hpc-compose runtime extensions, available at the top level and per service. See x-slurm.

Keyboard shortcuts

hpc-compose

Glossary

Related Docs