elasticgrid 0.1.0__py3-none-any.whl

elasticgrid/__init__.py

@@ -0,0 +1,40 @@
+"""elasticgrid: a cross-cloud elastic training grid.
+
+elasticgrid decides **how big** a training job should be right now and
+**where** those nodes should run. It is the control plane in front of
+iroh-tinker's multi-cloud launch layer:
+
+* :mod:`elasticgrid.metrics` collects the signals (step time, queue time,
+  throughput) from a rolling window of training steps;
+* :mod:`elasticgrid.scheduler` turns those signals into a scale decision,
+  with bounds, a cooldown and a deadline rule;
+* :mod:`elasticgrid.grid` places the nodes over the available providers and
+  costs the placement, then renders an iroh-tinker ``LaunchSpec``.
+
+Nothing here provisions anything by itself: the placement is a plan, and
+iroh-tinker's ``JobManager`` executes it.
+"""
+from __future__ import annotations
+
+__version__ = "0.1.0"
+
+from .grid import GridPlanner, Placement, Provider, build_launch_spec, providers_from_dicts
+from .metrics import MetricsWindow, StepSample
+from .scheduler import ElasticScheduler, ScaleAction, ScaleDecision
+
+__all__ = [
+    "__version__",
+    # metrics
+    "MetricsWindow",
+    "StepSample",
+    # scheduler
+    "ElasticScheduler",
+    "ScaleAction",
+    "ScaleDecision",
+    # grid
+    "GridPlanner",
+    "Placement",
+    "Provider",
+    "build_launch_spec",
+    "providers_from_dicts",
+]

elasticgrid/cli.py

@@ -0,0 +1,222 @@
+"""``elasticgrid`` command line interface.
+
+Subcommands:
+
+``providers``  show the configured grid
+``plan``       place N nodes over the grid and cost it
+``spec``       render the iroh-tinker launch spec for a placement
+``simulate``   feed a window of step samples through the scheduler
+
+Every command supports ``--json``.
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+from collections.abc import Sequence
+from pathlib import Path
+from typing import Any
+
+from .grid import GridPlanner, Provider, build_launch_spec, providers_from_dicts
+from .metrics import MetricsWindow, StepSample
+from .scheduler import ElasticScheduler
+
+__all__ = ["build_parser", "main"]
+
+DEFAULT_PROVIDERS = [
+    {
+        "name": "local",
+        "region": "",
+        "max_nodes": 8,
+        "cost_per_node_hour": 0.0,
+        "spot": False,
+    },
+    {
+        "name": "gcp",
+        "region": "us-central1",
+        "max_nodes": 64,
+        "cost_per_node_hour": 3.5,
+        "accelerator": "A100",
+        "spot": True,
+    },
+    {
+        "name": "aws",
+        "region": "us-east-1",
+        "max_nodes": 64,
+        "cost_per_node_hour": 3.2,
+        "accelerator": "p5.48xlarge",
+        "spot": True,
+    },
+    {
+        "name": "azure",
+        "region": "westeurope",
+        "max_nodes": 32,
+        "cost_per_node_hour": 3.9,
+        "accelerator": "ND96asr",
+        "spot": True,
+    },
+]
+
+
+def _load_providers(path: str | None) -> list[Provider]:
+    if path is None:
+        return providers_from_dicts(DEFAULT_PROVIDERS)
+    raw = json.loads(Path(path).read_text())
+    entries = raw.get("providers", raw) if isinstance(raw, dict) else raw
+    return providers_from_dicts(entries)
+
+
+def build_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(
+        prog="elasticgrid",
+        description="Cross-cloud elastic training grid for variable-size workloads.",
+    )
+    parser.add_argument("--json", action="store_true", help="emit JSON on stdout")
+    sub = parser.add_subparsers(dest="command", required=True)
+
+    providers = sub.add_parser("providers", help="show the configured grid")
+    providers.add_argument("--providers", help="JSON file describing the providers")
+    providers.add_argument("--json", action="store_true", help="emit JSON on stdout")
+
+    plan = sub.add_parser("plan", help="place nodes over the grid")
+    plan.add_argument("--nodes", type=int, required=True, help="how many nodes")
+    plan.add_argument("--providers", help="JSON file describing the providers")
+    plan.add_argument("--max-spot-fraction", type=float, default=1.0)
+    plan.add_argument("--prefer", default="cost", choices=["cost", "balance"])
+    plan.add_argument("--json", action="store_true", help="emit JSON on stdout")
+
+    spec = sub.add_parser("spec", help="render a launch spec for a placement")
+    spec.add_argument("--nodes", type=int, required=True, help="how many nodes")
+    spec.add_argument("--providers", help="JSON file describing the providers")
+    spec.add_argument("--name", default="grid-job")
+    spec.add_argument("--run", default="python train.py")
+    spec.add_argument("--json", action="store_true", help="emit JSON on stdout")
+
+    simulate = sub.add_parser("simulate", help="run the scheduler over a window")
+    simulate.add_argument(
+        "--samples",
+        required=True,
+        help="JSON list of {step, duration_s, queue_s, nodes}",
+    )
+    simulate.add_argument("--nodes", type=int, default=8, help="current node count")
+    simulate.add_argument("--min-nodes", type=int, default=1)
+    simulate.add_argument("--max-nodes", type=int, default=1024)
+    simulate.add_argument("--cooldown", type=float, default=300.0)
+    simulate.add_argument("--json", action="store_true", help="emit JSON on stdout")
+
+    return parser
+
+
+def _cmd_providers(args: argparse.Namespace) -> int:
+    planner = GridPlanner(_load_providers(args.providers))
+    payload = planner.summary()
+    if args.json:
+        _emit_json(payload)
+        return 0
+    print("provider  region        nodes  $/node/h  spot")
+    for provider in payload["providers"]:
+        print(
+            f"  {provider['name']:<8} {provider['region'] or '-':<12} "
+            f"{provider['max_nodes']:>5}  {provider['cost_per_node_hour']:>9.2f}  "
+            f"{provider['spot']}"
+        )
+    print(f"total capacity: {payload['capacity']} nodes")
+    return 0
+
+
+def _cmd_plan(args: argparse.Namespace) -> int:
+    planner = GridPlanner(
+        _load_providers(args.providers),
+        max_spot_fraction=args.max_spot_fraction,
+        prefer=args.prefer,
+    )
+    placement = planner.plan(args.nodes)
+    payload = placement.to_dict()
+    if args.json:
+        _emit_json(payload)
+        return 0
+    print(f"placement for {placement.total_nodes} nodes:")
+    for name, count in placement.assignments:
+        print(f"  {name:<10} {count:>5} nodes")
+    print(f"cost: ${placement.cost_per_hour:.2f}/hour  spot share: {placement.spot_fraction:.0%}")
+    print(f"infra order: {', '.join(placement.infra_refs)}")
+    return 0
+
+
+def _cmd_spec(args: argparse.Namespace) -> int:
+    planner = GridPlanner(_load_providers(args.providers))
+    placement = planner.plan(args.nodes)
+    spec = build_launch_spec(
+        placement,
+        name=args.name,
+        run=args.run,
+        num_nodes=placement.total_nodes,
+    )
+    payload = spec.to_dict()
+    if args.json:
+        _emit_json(payload)
+        return 0
+    print(json.dumps(payload, indent=2))
+    return 0
+
+
+def _cmd_simulate(args: argparse.Namespace) -> int:
+    raw = json.loads(Path(args.samples).read_text())
+    window = MetricsWindow()
+    for entry in raw:
+        window.add(
+            StepSample(
+                step=int(entry.get("step", 0)),
+                duration_s=float(entry.get("duration_s", 0.0)),
+                queue_s=float(entry.get("queue_s", 0.0)),
+                nodes=int(entry.get("nodes", args.nodes)),
+                tokens=int(entry.get("tokens", 0)),
+            )
+        )
+    scheduler = ElasticScheduler(
+        min_nodes=args.min_nodes,
+        max_nodes=args.max_nodes,
+        cooldown_s=args.cooldown,
+    )
+    decision = scheduler.decide(current_nodes=args.nodes, window=window)
+    payload = {
+        "window": window.to_dict(),
+        "decision": decision.to_dict(),
+    }
+    if args.json:
+        _emit_json(payload)
+        return 0
+    print(f"action: {decision.action} ({decision.current_nodes} -> {decision.target_nodes})")
+    print(f"reason: {decision.reason}")
+    return 0
+
+
+def _emit_json(payload: Any) -> None:
+    sys.stdout.write(json.dumps(payload, indent=2) + "\n")
+
+
+def main(argv: Sequence[str] | None = None) -> int:
+    parser = build_parser()
+    args = parser.parse_args(list(argv) if argv is not None else None)
+    try:
+        if args.command == "providers":
+            return _cmd_providers(args)
+        if args.command == "plan":
+            return _cmd_plan(args)
+        if args.command == "spec":
+            return _cmd_spec(args)
+        if args.command == "simulate":
+            return _cmd_simulate(args)
+    except (FileNotFoundError, ValueError, RuntimeError) as exc:
+        if getattr(args, "json", False):
+            _emit_json({"error": str(exc)})
+        else:
+            print(f"error: {exc}", file=sys.stderr)
+        return 2
+    parser.error(f"unknown command {args.command!r}")
+    return 2
+
+
+if __name__ == "__main__":  # pragma: no cover
+    raise SystemExit(main())

elasticgrid/grid.py

@@ -0,0 +1,304 @@
+"""Where the nodes go: the cross-cloud placement.
+
+The grid spreads a workload over several providers. The placement is a pure
+function of (nodes, providers, policy), so it can be planned, costed and
+reviewed before anything is provisioned, and two runs of the same plan get
+the same placement.
+
+The result is handed to iroh-tinker's launch layer as
+``ResourcesSpec.infra``: an ordered list of infra references such as
+``gcp/us-central1``. iroh-tinker fails over down that list, so the order
+matters: the cheapest provider that can take the whole job goes first, and
+the spot/on-demand mix is explicit.
+"""
+from __future__ import annotations
+
+from collections.abc import Iterable, Mapping, Sequence
+from dataclasses import dataclass
+from typing import Any
+
+from iroh_tinker.cloud.spec import LaunchSpec, ResourcesSpec
+
+__all__ = [
+    "GridPlanner",
+    "Placement",
+    "Provider",
+    "build_launch_spec",
+    "providers_from_dicts",
+]
+
+
+@dataclass(frozen=True)
+class Provider:
+    """A place to run nodes."""
+
+    name: str
+    region: str = ""
+    max_nodes: int = 0
+    cost_per_node_hour: float = 0.0
+    accelerator: str | None = None
+    spot: bool = True
+
+    def __post_init__(self) -> None:
+        if not self.name:
+            raise ValueError("provider name must not be empty")
+        if self.max_nodes < 0:
+            raise ValueError("max_nodes must be non-negative")
+        if self.cost_per_node_hour < 0:
+            raise ValueError("cost_per_node_hour must be non-negative")
+
+    @property
+    def infra_ref(self) -> str:
+        """The infra reference iroh-tinker's launch layer understands."""
+        return f"{self.name}/{self.region}" if self.region else self.name
+
+    @property
+    def capacity(self) -> int:
+        """How many nodes this provider can take right now."""
+        return self.max_nodes
+
+    def to_dict(self) -> dict[str, Any]:
+        return {
+            "name": self.name,
+            "region": self.region,
+            "max_nodes": self.max_nodes,
+            "cost_per_node_hour": self.cost_per_node_hour,
+            "accelerator": self.accelerator,
+            "spot": self.spot,
+            "infra_ref": self.infra_ref,
+        }
+
+
+@dataclass(frozen=True)
+class Placement:
+    """How many nodes each provider takes."""
+
+    assignments: tuple[tuple[str, int], ...]
+    total_nodes: int
+    cost_per_hour: float
+    providers: tuple[Provider, ...] = ()
+
+    def nodes_for(self, provider_name: str) -> int:
+        for name, count in self.assignments:
+            if name == provider_name:
+                return count
+        return 0
+
+    @property
+    def infra_refs(self) -> list[str]:
+        """The ordered infra references for the launch spec."""
+        by_name = {provider.name: provider for provider in self.providers}
+        refs = []
+        for name, count in self.assignments:
+            if count <= 0:
+                continue
+            provider = by_name.get(name)
+            if provider is not None:
+                refs.append(provider.infra_ref)
+        return refs
+
+    @property
+    def spot_fraction(self) -> float:
+        """Share of the nodes that are on spot capacity."""
+        if self.total_nodes <= 0:
+            return 0.0
+        by_name = {provider.name: provider for provider in self.providers}
+        spot_nodes = sum(
+            count
+            for name, count in self.assignments
+            if by_name.get(name) is not None and by_name[name].spot
+        )
+        return spot_nodes / self.total_nodes
+
+    def to_dict(self) -> dict[str, Any]:
+        return {
+            "assignments": [
+                {"provider": name, "nodes": count} for name, count in self.assignments
+            ],
+            "total_nodes": self.total_nodes,
+            "cost_per_hour": self.cost_per_hour,
+            "spot_fraction": self.spot_fraction,
+            "infra_refs": self.infra_refs,
+        }
+
+
+class GridPlanner:
+    """Plans a placement of N nodes over a set of providers."""
+
+    def __init__(
+        self,
+        providers: Sequence[Provider],
+        *,
+        max_spot_fraction: float = 1.0,
+        prefer: str = "cost",
+    ) -> None:
+        if not providers:
+            raise ValueError("a grid needs at least one provider")
+        if not 0 <= max_spot_fraction <= 1:
+            raise ValueError("max_spot_fraction must be in [0, 1]")
+        if prefer not in ("cost", "balance"):
+            raise ValueError("prefer must be 'cost' or 'balance'")
+        names = [provider.name for provider in providers]
+        if len(set(names)) != len(names):
+            raise ValueError("provider names must be unique")
+        self.providers = tuple(providers)
+        self.max_spot_fraction = max_spot_fraction
+        self.prefer = prefer
+
+    # -- ordering --------------------------------------------------------
+    def ordered(self) -> list[Provider]:
+        """The providers in the order the placement will use them."""
+        if self.prefer == "cost":
+            return sorted(self.providers, key=lambda p: (p.cost_per_node_hour, p.name))
+        # "balance": spread across providers, cheapest first.
+        return sorted(self.providers, key=lambda p: (p.cost_per_node_hour, p.name))
+
+    # -- planning ---------------------------------------------------------
+    def plan(self, nodes: int) -> Placement:
+        """Plan ``nodes`` nodes over the providers.
+
+        Raises ``ValueError`` when the grid cannot hold the requested size.
+        """
+        if nodes < 0:
+            raise ValueError("nodes must be non-negative")
+        if nodes == 0:
+            return Placement(
+                assignments=(),
+                total_nodes=0,
+                cost_per_hour=0.0,
+                providers=self.providers,
+            )
+
+        capacity = sum(provider.capacity for provider in self.providers)
+        if nodes > capacity:
+            raise ValueError(
+                f"the grid can hold {capacity} nodes, {nodes} requested"
+            )
+
+        order = self.ordered()
+        spot_budget = int(self.max_spot_fraction * nodes)
+        spot_used = 0
+        on_demand_used = 0
+        assignments: dict[str, int] = {provider.name: 0 for provider in order}
+
+        # First pass: fill the cheapest provider, but respect the spot budget.
+        for provider in order:
+            if sum(assignments.values()) >= nodes:
+                break
+            remaining = nodes - sum(assignments.values())
+            take = min(provider.capacity, remaining)
+            if provider.spot:
+                allowed = max(0, spot_budget - spot_used)
+                take = min(take, allowed)
+            else:
+                take = min(take, max(0, nodes - spot_budget - on_demand_used))
+            if take <= 0:
+                continue
+            assignments[provider.name] = take
+            if provider.spot:
+                spot_used += take
+            else:
+                on_demand_used += take
+
+        # Second pass: if the spot budget left nodes unplaced, fill them from
+        # the remaining capacity regardless of the spot policy, because a job
+        # that cannot start is worse than a job that breaks the policy.
+        shortfall = nodes - sum(assignments.values())
+        if shortfall > 0:
+            for provider in order:
+                if shortfall <= 0:
+                    break
+                take = min(shortfall, provider.capacity - assignments[provider.name])
+                if take <= 0:
+                    continue
+                assignments[provider.name] += take
+                shortfall -= take
+
+        ordered_assignments = tuple(
+            (provider.name, assignments[provider.name])
+            for provider in order
+            if assignments[provider.name] > 0
+        )
+        cost = sum(
+            count * provider.cost_per_node_hour
+            for provider in order
+            for name, count in ordered_assignments
+            if name == provider.name
+        )
+        return Placement(
+            assignments=ordered_assignments,
+            total_nodes=sum(count for _, count in ordered_assignments),
+            cost_per_hour=round(cost, 6),
+            providers=self.providers,
+        )
+
+    def plan_many(self, sizes: Iterable[int]) -> list[Placement]:
+        return [self.plan(size) for size in sizes]
+
+    def summary(self) -> dict[str, Any]:
+        return {
+            "providers": [provider.to_dict() for provider in self.ordered()],
+            "capacity": sum(provider.capacity for provider in self.providers),
+            "max_spot_fraction": self.max_spot_fraction,
+            "prefer": self.prefer,
+        }
+
+
+def build_launch_spec(
+    placement: Placement,
+    *,
+    name: str,
+    run: str,
+    workdir: str = ".",
+    num_nodes: int | None = None,
+    procs_per_node: int = 1,
+    backend: str = "numpy",
+    instance_type: str | None = None,
+    use_spot: bool = True,
+    rendezvous_uri: str | None = None,
+) -> LaunchSpec:
+    """Turn a placement into an iroh-tinker launch spec."""
+    nodes = placement.total_nodes if num_nodes is None else num_nodes
+    spec = LaunchSpec(
+        name=name,
+        run=run,
+        workdir=workdir,
+        num_nodes=nodes,
+        procs_per_node=procs_per_node,
+        backend=backend,
+        resources=ResourcesSpec(
+            infra=placement.infra_refs,
+            instance_type=instance_type,
+            use_spot=use_spot,
+        ),
+        rendezvous=_rendezvous(rendezvous_uri),
+    )
+    spec.validate()
+    return spec
+
+
+def _rendezvous(uri: str | None) -> Any:
+    from iroh_tinker.cloud.spec import RendezvousSpec
+
+    if not uri:
+        return RendezvousSpec()
+    if uri.startswith("dir:"):
+        return RendezvousSpec(mode="dir", uri=uri[len("dir:") :])
+    return RendezvousSpec(mode="dir", uri=uri)
+
+
+def providers_from_dicts(raw: Iterable[Mapping[str, Any]]) -> list[Provider]:
+    """Build providers from a JSON-ish description."""
+    providers: list[Provider] = []
+    for entry in raw:
+        providers.append(
+            Provider(
+                name=str(entry["name"]),
+                region=str(entry.get("region", "")),
+                max_nodes=int(entry.get("max_nodes", 0)),
+                cost_per_node_hour=float(entry.get("cost_per_node_hour", 0.0)),
+                accelerator=entry.get("accelerator"),
+                spot=bool(entry.get("spot", True)),
+            )
+        )
+    return providers

elasticgrid/metrics.py

@@ -0,0 +1,145 @@
+"""Signals the elastic scheduler acts on.
+
+The scheduler does not look at the cloud; it looks at a window of step
+samples. Each sample says how long a step took, how long the step waited in
+the queue, how many nodes were running and how much work the step did.
+
+From those the window derives the two numbers that matter:
+
+* **queue time** -- workers are idle while they wait. That is a capacity
+  problem: add workers.
+* **step time** -- workers are busy but the step is slow. That is a
+  parallelism problem: either the step is too big for the current world size,
+  or the interconnect is the bottleneck.
+"""
+from __future__ import annotations
+
+from collections import deque
+from dataclasses import dataclass
+from typing import Any
+
+__all__ = ["MetricsWindow", "StepSample"]
+
+
+@dataclass(frozen=True)
+class StepSample:
+    """One observed training step."""
+
+    step: int
+    duration_s: float
+    queue_s: float = 0.0
+    nodes: int = 1
+    tokens: int = 0
+
+    def __post_init__(self) -> None:
+        if self.duration_s < 0:
+            raise ValueError("duration_s must be non-negative")
+        if self.queue_s < 0:
+            raise ValueError("queue_s must be non-negative")
+        if self.nodes < 0:
+            raise ValueError("nodes must be non-negative")
+        if self.tokens < 0:
+            raise ValueError("tokens must be non-negative")
+
+    @property
+    def active_s(self) -> float:
+        """The part of the step that was actually computing."""
+        return max(0.0, self.duration_s - self.queue_s)
+
+
+class MetricsWindow:
+    """A rolling window of step samples.
+
+    The window is the only thing the scheduler reads, so a decision can be
+    replayed from a recorded window in a test or an incident review.
+    """
+
+    def __init__(self, size: int = 10) -> None:
+        if size <= 0:
+            raise ValueError("window size must be positive")
+        self.size = size
+        self._samples: deque[StepSample] = deque(maxlen=size)
+
+    def add(self, sample: StepSample) -> None:
+        self._samples.append(sample)
+
+    def extend(self, samples: list[StepSample]) -> None:
+        for sample in samples:
+            self.add(sample)
+
+    def clear(self) -> None:
+        self._samples.clear()
+
+    # -- contents --------------------------------------------------------
+    def __len__(self) -> int:
+        return len(self._samples)
+
+    def __iter__(self):  # type: ignore[no-untyped-def]
+        return iter(self._samples)
+
+    @property
+    def samples(self) -> list[StepSample]:
+        return list(self._samples)
+
+    @property
+    def nodes(self) -> int:
+        """The node count of the most recent sample."""
+        return self._samples[-1].nodes if self._samples else 0
+
+    # -- derived signals -------------------------------------------------
+    @property
+    def mean_step_s(self) -> float:
+        if not self._samples:
+            return 0.0
+        return sum(s.duration_s for s in self._samples) / len(self._samples)
+
+    @property
+    def mean_queue_s(self) -> float:
+        if not self._samples:
+            return 0.0
+        return sum(s.queue_s for s in self._samples) / len(self._samples)
+
+    @property
+    def mean_active_s(self) -> float:
+        if not self._samples:
+            return 0.0
+        return sum(s.active_s for s in self._samples) / len(self._samples)
+
+    @property
+    def queue_fraction(self) -> float:
+        """Share of the step spent waiting rather than computing."""
+        total = sum(s.duration_s for s in self._samples)
+        if total <= 0:
+            return 0.0
+        return sum(s.queue_s for s in self._samples) / total
+
+    @property
+    def throughput_tokens_per_s(self) -> float:
+        if not self._samples:
+            return 0.0
+        tokens = sum(s.tokens for s in self._samples)
+        seconds = sum(s.duration_s for s in self._samples)
+        if seconds <= 0:
+            return 0.0
+        return tokens / seconds
+
+    @property
+    def saturated(self) -> bool:
+        """Whether the window shows the workers are saturated."""
+        return self.queue_fraction >= 0.2
+
+    @property
+    def stalled(self) -> bool:
+        """Whether the window shows the step is dominated by waiting."""
+        return self.queue_fraction >= 0.6
+
+    def to_dict(self) -> dict[str, Any]:
+        return {
+            "samples": len(self._samples),
+            "nodes": self.nodes,
+            "mean_step_s": self.mean_step_s,
+            "mean_queue_s": self.mean_queue_s,
+            "mean_active_s": self.mean_active_s,
+            "queue_fraction": self.queue_fraction,
+            "throughput_tokens_per_s": self.throughput_tokens_per_s,
+        }

elasticgrid/scheduler.py

@@ -0,0 +1,231 @@
+"""The elastic scale decision.
+
+The scheduler answers one question: given what the last few steps looked
+like, how many nodes should the next step run on?
+
+The rules, in order:
+
+1. **Never scale on thin evidence.** Fewer than ``min_samples`` observations
+   means hold.
+2. **Respect the bounds.** The target is clamped to
+   ``[min_nodes, max_nodes]``.
+3. **Honour the cooldown.** After a change, wait ``cooldown_s`` before
+   changing again. Elastic scaling without a cooldown oscillates.
+4. **Queue-bound means scale up.** If the step spent most of its time waiting
+   for a worker, the job needs more workers.
+5. **Idle-bound means scale down.** If the step was fast and nobody waited,
+   the job is paying for idle workers.
+6. **Do not scale up near the deadline.** Starting workers that will not
+   finish before the deadline costs money and saves nothing.
+"""
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import Any
+
+from .metrics import MetricsWindow
+
+__all__ = ["ElasticScheduler", "ScaleAction", "ScaleDecision"]
+
+
+class ScaleAction:
+    """The three things the scheduler can say."""
+
+    SCALE_UP = "scale_up"
+    SCALE_DOWN = "scale_down"
+    HOLD = "hold"
+
+    ALL = (SCALE_UP, SCALE_DOWN, HOLD)
+
+
+@dataclass(frozen=True)
+class ScaleDecision:
+    """One decision, with the reason it was made."""
+
+    action: str
+    current_nodes: int
+    target_nodes: int
+    reason: str
+    signals: dict[str, Any] = field(default_factory=dict)
+
+    @property
+    def changed(self) -> bool:
+        return self.action != ScaleAction.HOLD
+
+    def to_dict(self) -> dict[str, Any]:
+        return {
+            "action": self.action,
+            "current_nodes": self.current_nodes,
+            "target_nodes": self.target_nodes,
+            "reason": self.reason,
+            "signals": dict(self.signals),
+        }
+
+
+class ElasticScheduler:
+    """Decides the node count for the next step."""
+
+    def __init__(
+        self,
+        *,
+        min_nodes: int = 1,
+        max_nodes: int = 1024,
+        scale_up_factor: float = 2.0,
+        scale_down_factor: float = 0.5,
+        cooldown_s: float = 300.0,
+        target_queue_fraction: float = 0.2,
+        idle_queue_fraction: float = 0.05,
+        idle_step_fraction: float = 0.5,
+        target_step_s: float = 30.0,
+        min_samples: int = 3,
+        deadline_s: float | None = None,
+    ) -> None:
+        if min_nodes < 1:
+            raise ValueError("min_nodes must be at least 1")
+        if max_nodes < min_nodes:
+            raise ValueError("max_nodes must be >= min_nodes")
+        if scale_up_factor <= 1:
+            raise ValueError("scale_up_factor must be > 1")
+        if not 0 < scale_down_factor < 1:
+            raise ValueError("scale_down_factor must be in (0, 1)")
+        if min_samples < 1:
+            raise ValueError("min_samples must be at least 1")
+        self.min_nodes = min_nodes
+        self.max_nodes = max_nodes
+        self.scale_up_factor = scale_up_factor
+        self.scale_down_factor = scale_down_factor
+        self.cooldown_s = cooldown_s
+        self.target_queue_fraction = target_queue_fraction
+        self.idle_queue_fraction = idle_queue_fraction
+        self.idle_step_fraction = idle_step_fraction
+        self.target_step_s = target_step_s
+        self.min_samples = min_samples
+        self.deadline_s = deadline_s
+        self._last_change_at: float | None = None
+
+    # -- state ------------------------------------------------------------
+    def reset(self) -> None:
+        """Forget the cooldown (used when a new job starts)."""
+        self._last_change_at = None
+
+    def _clamp(self, nodes: int) -> int:
+        return max(self.min_nodes, min(self.max_nodes, nodes))
+
+    # -- decision ---------------------------------------------------------
+    def decide(
+        self,
+        *,
+        current_nodes: int,
+        window: MetricsWindow,
+        now: float = 0.0,
+        time_remaining_s: float | None = None,
+    ) -> ScaleDecision:
+        """Decide the node count for the next step."""
+        signals = {
+            "samples": len(window),
+            "queue_fraction": window.queue_fraction,
+            "mean_step_s": window.mean_step_s,
+            "mean_queue_s": window.mean_queue_s,
+            "nodes": window.nodes or current_nodes,
+            "cooldown_s": self.cooldown_s,
+        }
+
+        if len(window) < self.min_samples:
+            return ScaleDecision(
+                action=ScaleAction.HOLD,
+                current_nodes=current_nodes,
+                target_nodes=current_nodes,
+                reason="not enough samples to decide",
+                signals=signals,
+            )
+
+        if (
+            self._last_change_at is not None
+            and now - self._last_change_at < self.cooldown_s
+        ):
+            return ScaleDecision(
+                action=ScaleAction.HOLD,
+                current_nodes=current_nodes,
+                target_nodes=current_nodes,
+                reason="cooldown",
+                signals=signals,
+            )
+
+        if window.queue_fraction >= self.target_queue_fraction:
+            target = self._clamp(int(current_nodes * self.scale_up_factor))
+            if target == current_nodes:
+                return ScaleDecision(
+                    action=ScaleAction.HOLD,
+                    current_nodes=current_nodes,
+                    target_nodes=current_nodes,
+                    reason="already at the node ceiling",
+                    signals=signals,
+                )
+            if time_remaining_s is not None and time_remaining_s <= 0:
+                return ScaleDecision(
+                    action=ScaleAction.HOLD,
+                    current_nodes=current_nodes,
+                    target_nodes=current_nodes,
+                    reason="deadline reached, no time to start new workers",
+                    signals=signals,
+                )
+            reason = (
+                f"{window.queue_fraction:.0%} of the step was spent waiting for a "
+                f"worker; scaling up {current_nodes} -> {target}"
+            )
+            self._last_change_at = now
+            return ScaleDecision(
+                action=ScaleAction.SCALE_UP,
+                current_nodes=current_nodes,
+                target_nodes=target,
+                reason=reason,
+                signals=signals,
+            )
+
+        if (
+            window.mean_step_s > 0
+            and window.mean_step_s <= self.idle_step_fraction * self.target_step_s
+            and window.queue_fraction <= self.idle_queue_fraction
+        ):
+            # The job is comfortably faster than the target and nobody is
+            # waiting for a worker: the current size is more than enough.
+            if current_nodes <= self.min_nodes:
+                return ScaleDecision(
+                    action=ScaleAction.HOLD,
+                    current_nodes=current_nodes,
+                    target_nodes=current_nodes,
+                    reason="already at the node floor",
+                    signals=signals,
+                )
+            target = self._clamp(
+                max(self.min_nodes, int(current_nodes * self.scale_down_factor))
+            )
+            if target == current_nodes:
+                return ScaleDecision(
+                    action=ScaleAction.HOLD,
+                    current_nodes=current_nodes,
+                    target_nodes=current_nodes,
+                    reason="already at the node floor",
+                    signals=signals,
+                )
+            reason = (
+                f"the step takes {window.mean_step_s:.1f}s, well inside the "
+                f"{self.target_step_s:.1f}s target, and no worker is waiting; "
+                f"scaling down {current_nodes} -> {target}"
+            )
+            self._last_change_at = now
+            return ScaleDecision(
+                action=ScaleAction.SCALE_DOWN,
+                current_nodes=current_nodes,
+                target_nodes=target,
+                reason=reason,
+                signals=signals,
+            )
+
+        return ScaleDecision(
+            action=ScaleAction.HOLD,
+            current_nodes=current_nodes,
+            target_nodes=current_nodes,
+            reason="the step is balanced; hold the current size",
+            signals=signals,
+        )

elasticgrid-0.1.0.dist-info/METADATA

@@ -0,0 +1,146 @@
+Metadata-Version: 2.4
+Name: elasticgrid
+Version: 0.1.0
+Summary: Cross-cloud elastic training grid for variable-size workloads
+License: Apache-2.0
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: Apache Software License
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Programming Language :: Python :: 3.14
+Classifier: Programming Language :: Python :: 3.15
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: iroh-tinker>=0.3.0
+Provides-Extra: dev
+Requires-Dist: pytest>=8.0; extra == "dev"
+Requires-Dist: pytest-asyncio>=0.23; extra == "dev"
+Requires-Dist: ruff>=0.5; extra == "dev"
+Requires-Dist: mypy>=1.10; extra == "dev"
+Dynamic: license-file
+
+# elasticgrid
+
+Cross-cloud elastic training grid for variable-size workloads, built on
+[iroh-tinker](https://github.com/scairesearch/iroh_tinker).
+
+elasticgrid is the control plane in front of iroh-tinker's multi-cloud launch
+layer. It answers two questions:
+
+* **How big should this job be right now?** A rolling window of training
+  steps (step time, queue time, throughput) feeds a scheduler that grows the
+  job when workers are waiting and shrinks it when the job is comfortably
+  faster than the target, with bounds, a cooldown and a deadline rule.
+* **Where should the nodes run?** A placement plan spreads the nodes over the
+  available providers, respects a spot/on-demand budget, costs the placement
+  and renders it as an iroh-tinker `LaunchSpec`.
+
+Nothing here provisions anything by itself: the placement is a plan, and
+iroh-tinker's `JobManager` executes it.
+
+## Install
+
+```bash
+pip install elasticgrid
+pip install -e ".[dev]"           # from a checkout
+```
+
+elasticgrid depends on the public `iroh-tinker` surface only.
+
+## Built on the iroh-tinker public surface
+
+| API | Used for |
+|---|---|
+| `iroh_tinker.cloud.spec.LaunchSpec` / `ResourcesSpec` / `RendezvousSpec` | the launch spec a placement renders into |
+| `iroh_tinker.cloud.job.JobManager` | executing the plan |
+| `iroh_tinker.hw.HardwareContext` / `infer_topology` / `select_topology_params` | topology hints for the placement |
+
+`iroh_tinker.AllreduceAlgorithm` is **not** exported from the top-level
+package; it lives in `iroh_tinker.comm`. elasticgrid does not depend on it.
+
+## CLI
+
+```bash
+elasticgrid providers                        # the configured grid
+elasticgrid plan --nodes 64                 # placement + cost
+elasticgrid plan --nodes 64 --max-spot-fraction 0.5
+elasticgrid spec --nodes 64 --name my-job   # the iroh-tinker launch spec
+elasticgrid simulate --samples steps.json   # what the scheduler would do
+```
+
+Every command accepts `--json`.
+
+## The scale decision
+
+The scheduler reads a window of step samples and applies the rules in order:
+
+1. **Never scale on thin evidence.** Fewer than `min_samples` observations
+   means hold.
+2. **Respect the bounds.** The target is clamped to `[min_nodes, max_nodes]`.
+3. **Honour the cooldown.** After a change, wait `cooldown_s`. Elastic
+   scaling without a cooldown oscillates.
+4. **Queue-bound means scale up.** If 20% or more of the step was spent
+   waiting for a worker, the job needs more workers.
+5. **Idle means scale down.** If the step is well inside `target_step_s` and
+   nobody is waiting, the current size is more than enough.
+6. **Do not scale up past the deadline.** Workers that cannot finish before
+   the deadline cost money and save nothing.
+
+Every decision carries the signals it was based on, so it can be replayed and
+reviewed:
+
+```json
+{
+  "action": "scale_up",
+  "current_nodes": 8,
+  "target_nodes": 16,
+  "reason": "80% of the step was spent waiting for a worker; scaling up 8 -> 16",
+  "signals": {"queue_fraction": 0.8, "mean_step_s": 10.0, "...": "..."}
+}
+```
+
+## The placement
+
+`GridPlanner.plan(nodes)` returns a placement: how many nodes each provider
+takes, what it costs per hour, and the ordered infra references that
+iroh-tinker fails over through.
+
+The policy is explicit:
+
+* providers are filled cheapest first;
+* `max_spot_fraction` caps the share of nodes on spot capacity;
+* if the spot budget cannot hold the job, the remaining nodes are placed
+  anyway, because a job that cannot start is worse than a job that breaks
+  the policy;
+* the placement is a pure function of (nodes, providers, policy), so the
+  same plan always gives the same answer.
+
+## Tests
+
+```bash
+pip install -e ".[dev]"
+pytest
+```
+
+The suite is credential-free: the planner and the scheduler are pure
+functions and are tested directly.
+
+## Layout
+
+```
+src/elasticgrid/
+  metrics.py   step samples and the rolling window of signals
+  scheduler.py the scale decision (bounds, cooldown, deadline)
+  grid.py      providers, placement, cost, launch spec
+  cli.py       elasticgrid providers | plan | spec | simulate
+```
+
+## Licence
+
+Apache-2.0. elasticgrid is a separate project from iroh-tinker, which stays
+MIT OR Apache-2.0.

elasticgrid-0.1.0.dist-info/RECORD

@@ -0,0 +1,12 @@
+elasticgrid/__init__.py,sha256=ZJSB1LoGLzRjG3O0dsLS4W0wiH1L23KiqhubciohvD0,1298
+elasticgrid/cli.py,sha256=YNeAsZAFtKFhGvTjVO5Tx3_6BIdIDWOgtDIlI_ACcTA,7476
+elasticgrid/grid.py,sha256=mvzlvN7xdOQHYgtSb21VAAs4HAhxN85Hh5-yI0YBkB0,10400
+elasticgrid/metrics.py,sha256=3lZGWzEbhk0SWwyD5Sla0v35g5MAhxCZ9JsPrZSEshw,4646
+elasticgrid/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+elasticgrid/scheduler.py,sha256=1xTs9lGBd_c4XCUBrDh-wWvBs9gqJX8XntVXL8t-jXc,8327
+elasticgrid-0.1.0.dist-info/licenses/LICENSE,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
+elasticgrid-0.1.0.dist-info/METADATA,sha256=inMRrKvnm3cUvXbpsoc146-Tj6QfmJYdx1TewjdgT-0,5175
+elasticgrid-0.1.0.dist-info/WHEEL,sha256=mOdXR9C3wMVov7cSY8PY-XCzo3SmmWCHvGQr5csJoao,91
+elasticgrid-0.1.0.dist-info/entry_points.txt,sha256=yfH5fSyfcN-mzRhzsUiPIQNVsLM2QyDCjz9tndiQuoM,53
+elasticgrid-0.1.0.dist-info/top_level.txt,sha256=KjyHnhFMAld5zHO7VgL2Wj4QhZ1SNS8sQEPYoGlyBD4,12
+elasticgrid-0.1.0.dist-info/RECORD,,

elasticgrid-0.1.0.dist-info/WHEEL

@@ -0,0 +1,5 @@
+Wheel-Version: 1.0
+Generator: setuptools (84.2.1)
+Root-Is-Purelib: true
+Tag: py3-none-any
+

elasticgrid-0.1.0.dist-info/entry_points.txt

@@ -0,0 +1,2 @@
+[console_scripts]
+elasticgrid = elasticgrid.cli:main

elasticgrid-0.1.0.dist-info/licenses/LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

elasticgrid-0.1.0.dist-info/top_level.txt

@@ -0,0 +1 @@
+elasticgrid