Servatus

Run resumable work through Slurm and atomically publish validated outputs.

Servatus 0.6.0 combines durable publication with the native Slurm Campaign interface below.

pip install servatus

Campaigns

A Campaign freezes an ordered prefix of opaque tasks. Reopening with the exact sequence is idempotent; reopening with that exact prefix plus a nonempty suffix durably registers the new tasks. Removing, reordering, or changing any registered task fails. Planning is local and deterministic. Submission records durable intent before contacting Slurm, records the acceptance receipt afterward, and stops on an ambiguous missing receipt rather than risking duplicate work.

from pathlib import Path, PurePosixPath

from servatus import Campaign, ResourceRequest, SlurmTarget, Task

campaign = Campaign.open(
    Path("state/training"),
    [Task("candidate-0", ("train", "--candidate", "0"), b'{"seed": 7}\n')],
)
resources = ResourceRequest(
    cpus_per_task=8,
    memory_mib_per_task=32768,
    gpus_per_task=1,
    time_limit="1-00:00:00",
)
target = SlurmTarget.from_toml(Path("TARGET.toml"))
plan = campaign.plan(target, resources)
# Review plan.allocations and plan.digest, then submit explicitly:
# validation = campaign.validate(plan)  # Optional, time-specific Slurm --test-only check.
receipts = campaign.submit(plan)
# Later processes reopen the durable roster without resupplying Tasks:
campaign = Campaign.load(Path("state/training"))

ResourceRequest has no defaults. CPU and MiB memory are positive, GPUs are a nonnegative whole count, and time uses canonical [days-]hours:minutes:seconds. One request applies to every Task in one Campaign. CPU-only, one-GPU, and one-process whole-multi-GPU tasks are supported. A project uses separate Campaigns for different resource shapes.

The request retains the authored wall time for provenance. Plans and sbatch show Slurm's effective limit, rounded upward once to whole minutes. The same rounding is applied before comparing a request with the target ceiling; packed task count never multiplies wall time.

RESOURCES.toml contains exactly four required values:

cpus_per_task = 32
memory_mib_per_task = 65536
gpus_per_task = 1
time_limit = "3-00:00:00"

TARGET.toml describes one concrete execution lane and its conservative guardrails:

host = "login.example.edu"
slurm_bin = "/opt/slurm/bin"
apptainer = "/usr/bin/apptainer"
image = "/cluster/images/project.sif"
work_root = "/cluster/work/project"
log_root = "/cluster/logs/project"
partitions = ["gpu"]
account = "research" # optional; qos and constraint are also optional
gpu_gres = "gpu"     # omit for a CPU-only target
max_tasks_per_allocation = 4
max_cpus_per_allocation = 128
max_memory_mib_per_allocation = 262144
max_gpus_per_allocation = 4 # use 0 when gpu_gres is omitted
max_time_limit = "7-00:00:00"
max_allocations_per_submit = 64
max_script_bytes = 4194304

Unknown keys, counted GRES, relative remote paths, unsafe site tokens, controls, booleans used as integers, unlimited/zero resources, and conflicting GPU settings are rejected. A target profile is a user-side mistake guard, not cluster authorization. Every listed partition must fit one truthful conservative envelope.

The planner preserves authored order and uses the fewest balanced groups allowed by every declared ceiling. An allocation containing n Tasks requests exactly n*C CPUs, n*M MiB, and n*G GPUs; time remains T. A caller may lower packing with tasks_per_allocation, but a cap above feasible capacity is rejected rather than clamped. Servatus never rounds up to node capacity.

Append-only growth preserves target/resource lineage, accepted receipts, retry history, and ambiguous intents. It increments campaign revision, so a plan made before the append becomes stale. Accepted prefix tasks are not selected again unless the caller explicitly requests retry.

The current development line uses Campaign and plan schema 3. Loading a plan regenerates it from typed inputs and requires identical canonical bytes. Schema 2 state and plans are rejected; create a new Campaign when upgrading.

Each allocation runs one concurrent srun --exclusive --exact --nodes=1 --ntasks=1 step per Task. Each step receives its exact CPU, MiB, and whole-GPU request and starts the target's immutable Apptainer image from work_root. CPU-only work emits no GRES or --nv. Servatus never emits job-level exclusivity, overlap, all memory, manual CUDA indices, ranks, or raw scheduler flags.

Slurm writes combined allocation stdout/stderr to log_root/%j.out and each combined task stream to log_root/%j-<zero-based-slot>.out. %j is expanded by Slurm after it assigns the job ID; plans and durable intent therefore remain immutable before scheduler acceptance.

Actual simultaneous placement depends on truthful resource requests and site CPU/GRES topology. Servatus does not silently inflate CPU requests, disable binding, or expose raw scheduler flags.

Task arguments and byte-exact stdin are embedded in the complete batch script before sbatch acceptance. They are excluded from ordinary plan and status output, but are not secrets: cluster administrators and accounting systems may be able to inspect them. Scheduler names expose only a random Servatus allocation identity.

CLI

The Python interface is authoritative. The CLI task JSONL adapter has exactly key, string-array args, and stdin_file per line:

servatus plan TASKS.jsonl --target TARGET.toml --resources RESOURCES.toml \
  --campaign STATE_DIR --output PLAN.json --tasks-per-allocation 4
# Exclude caller-validated work or explicitly retry accepted work; both flags are repeatable:
servatus plan TASKS.jsonl --target TARGET.toml --resources RESOURCES.toml \
  --campaign STATE_DIR --output RETRY.json --completed task-2 --retry task-0
# Explicit sensitive diagnostic; prints complete scripts, arguments, and payloads:
servatus plan TASKS.jsonl --target TARGET.toml --resources RESOURCES.toml \
  --campaign STATE_DIR --output PLAN.json --show-scripts
servatus validate STATE_DIR PLAN.json
servatus submit STATE_DIR PLAN.json
servatus status STATE_DIR
servatus reconcile STATE_DIR ALLOCATION_ID
servatus resolve STATE_DIR ALLOCATION_ID --job-id 1234 --cluster alpha
servatus resolve STATE_DIR ALLOCATION_ID --not-submitted

To extend an existing Campaign through the CLI, pass the complete previously registered JSONL prefix followed by the new suffix. Supplying only the suffix or changing the prefix fails closed. Relative stdin_file paths resolve against TASKS.jsonl's parent.

PLAN.json is published owner-only and never overwrites an existing path. It contains task keys, requested resources, effective allocation totals, target values, exact nonsecret sbatch arguments, allocation identities, and digests—not task arguments or stdin. Complete scripts are shown only by the warning-bearing --show-scripts diagnostic. validate makes one serial sbatch --test-only call per distinct allocation shape and prints each stable shape/script digest plus the controller response. Its answer is time-specific and does not submit or mutate campaign state.

An intent without a receipt is ambiguous. reconcile uses the Campaign's validated target lineage for one bounded squeue/sacct query and adopts only one exact Servatus identity. Otherwise an operator must resolve it explicitly as accepted or not submitted. Retry is explicit through Campaign.plan(..., retry={...}); prior receipts remain in history. Status reports unaccepted tasks, not incomplete work: the caller supplies completed after its own canonical validation.

Servatus does not cancel jobs in V1. Use the receipt with the site's normal scancel command. Cancellation applies to the packed allocation, does not prove application completion, and does not enable retry automatically.

Publication

Use publish when failed work is disposable:

from pathlib import Path

from servatus import Draft, publish


def build(draft: Draft) -> None:
    (draft.path / "result.json").write_text('{"status":"complete"}\n')


publication = publish(Path("outputs/run-1"), build)

The builder owns draft contents and validation. It must finish all content mutations before returning; Servatus syncs the quiescent draft afterward.

When the application has finished authoring an owner-only sibling tree, publish can retire that tree only after the canonical destination is committed and synced:

bundle = Path("outputs/.run-1.active")
publication = publish(Path("outputs/run-1"), build, retire=bundle)

The retained tree must already exist, differ from the destination, share its exact parent, and be quiescent before the call. Servatus pins it before entering the builder. Builder, validation, collision, and other precommit failures preserve it. After a durable commit, Servatus removes only the pinned tree and syncs the parent again. If that exact removal or its durability cannot be proved, publication still succeeds, Publication.cleanup_pending is true, and one best-effort RuntimeWarning reports the residue.

Use publish_file for one canonical regular file:

from pathlib import Path

from servatus import publish_file


def write(stage: Path) -> None:
    stage.write_text('{"status":"complete"}\n')
    # Perform application validation before returning.


publication = publish_file(Path("outputs/protocol.json"), write)

The writer receives an existing empty adjacent regular file. It must write and validate that inode in place; unlinking, replacing, or changing its file type fails publication. Its initial mode is created from 0o666 through the process umask, and an explicit writer chmod is preserved. The writer must finish all content mutations before returning.

Use Workspace when a worker must retain private checkpoints across restarts:

from servatus import Draft, Workspace

destination = Path("outputs/model-1")
with Workspace(destination, identity=b"model request bytes") as workspace:
    checkpoint = workspace.path / "last.ckpt"
    # The application creates or resumes its own checkpoint here.

    def assemble(draft: Draft) -> None:
        draft.link(checkpoint, "last.ckpt")
        # Perform application validation before returning.

    publication = workspace.publish(assemble)

Workspace binds stable hidden state to the SHA-256 digest of opaque identity bytes and holds a nonblocking writer lock. Draft.link atomically hard-links the source path into a safe relative path, then accepts only a same-filesystem regular file. The hard-link operation selects the source inode, so a safe source-path replacement before that operation may be selected. The builder owns contents, validation, schemas, and completion meaning and must finish mutating linked contents before returning. The owner-only draft namespace must remain quiescent during each Draft.link(); hostile same-account replacement of its destination leaf during that call is outside the contract.

Independent workers can publish resumable child results beneath one future destination without entering the parent:

parent = Workspace(Path("outputs/study-1"), identity=b"study request bytes")

with parent.child("method-0", identity=b"method request bytes") as child:
    checkpoint = child.path / "last.ckpt"
    # Create or resume application work, then retain one immutable child result.
    child.publish(lambda draft: draft.link(checkpoint, "result.bin"))

# After the application decides all required children are valid:
with parent as workspace:
    workspace.publish(
        lambda draft: draft.link(parent.path / "method-0/result.bin", "method-0/result.bin")
    )

child() accepts one safe leaf and opaque identity. Different children may run concurrently; the same child and parent finalization remain exclusive and nonblocking. A failed child retains only its resumable private work, while a published child becomes immutable input under the parent work. Servatus does not track expected children, readiness, dependencies, or application completion.

The owner-only hidden Workspace container is the lifecycle trust root. Servatus pins and rechecks its entries without following links; unsafe substitution is preserved and reported as pending cleanup. Callers must protect the parent directory, run only trusted same-account code, and use a filesystem with stable cross-client inode identities. See SECURITY.md.

Guarantees and support boundary

• Campaign files are bounded, owner-only, schema-versioned, symlink-safe, atomically replaced, and synced. Every durable read validates the complete snapshot, including the rule that one allocation cannot be both accepted and resolved as not submitted.
• Intent preserves the normalized route, guardrails, requested resources, exact allocation totals, and reviewed nonsecret sbatch command before external acceptance.
• A destination is absent or one complete regular file or directory. Native commits and the Linux regular-file fallback never overwrite an existing entry. An already-present destination is rejected before its callback; the commit remains no-replace against later races. The Linux directory fallback serializes cooperating Servatus publishers with an exclusive parent-directory lock.
• Work, hard-link sources, stages, and destination must share a filesystem.
• Disposable stage names are not synced merely by creation. Files and directories are synced before commit; the parent is synced after publication.
• Builders and writers must be quiescent when they return. Servatus verifies pathname/inode identity and syncs content, but does not detect or exclude concurrent content writers.
• Builder failures expose no destination. Resumable work remains; disposable stages are removed.
• Successful workspace publication exactly removes its pinned private tree. A moved, substituted, or unremovable tree remains visible as cleanup residue and is reported separately.
• Directory publication can retire one exact owner-only sibling after commit. Precommit failure preserves it; unsafe or incomplete post-commit retirement sets cleanup_pending without changing publication success.
• Child workspaces share the parent lifecycle lease; parent publication is busy until they close.

Publication supports POSIX filesystems on Linux and macOS. Its Linux fallbacks require cooperating Servatus publishers, coherent advisory locks, owner-controlled parents, and stable inode identities; hardware durability still depends on the filesystem and mount. Campaign submission is an unprivileged workstation-side OpenSSH client for homogeneous independent processes in one-node Slurm allocations. See SECURITY.md and ADR 0003 for the exact fallback and live-acceptance envelope.

Non-goals

Servatus is not an ML framework, scheduler plugin, daemon, security boundary, experiment tracker, DAG engine, secrets manager, or transfer/image-deployment tool. V1 has no Submitit or runtime Python dependency, plugin/backend abstraction, local executor, arrays, heterogeneous tasks, multi-node ranks, MPI/torchrun, fractional/shared GPUs, queue-aware packing, automatic retry, background polling, cancellation/requeue, raw Slurm/environment passthrough, application completion probes, compatibility shims, or cross-filesystem copy fallback.

See the context glossary and architecture decisions for the ownership boundary.