Jobs and clusters¶
A cluster is a named destination: which scheduler to use, how to reach it, and where to place job files. It may be your laptop or an HPC login account.
A job is one allocation. It contains one command or a plan of commands that run inside that allocation.
| Value | Describes | Example |
|---|---|---|
Resources |
Allocation size | 4 tasks, 2 CPUs per task, 8 GB per node |
Scheduling |
Placement in the queue | Account, partition, name, dependencies |
Execution |
Work inside the allocation | Commands, environment, working directory |
JobRef |
A saved allocation identity | Registry, cluster, native ID, incarnation |
An execution unit is one command. Sequence runs units in order; Parallel
runs branches together. Both compose into a single job. Use separate submissions
when you want separate allocations. See execution plans.
Where state lives¶
The scheduler owns queue and accounting records. For local Shell jobs, target launch and exit receipts provide execution evidence. The registry stores cluster definitions; it does not copy job history into a database.
A reference includes the native ID and its incarnation. Keep the whole reference: a bare Slurm or PBS number is insufficient to identify a molq job safely.
Connecting¶
Molq() starts a temporary service over stdio. A shared molq serve endpoint lets
the CLI, SDKs, and watcher use the same observations and event cursors. Each
independent service has its own cache and cursors.
Both connections use the same public API. See configuration for a shared service and monitoring for status freshness.