stroblmeandClaude Opus 5
3c15964364
Docs / docs (push) Successful in 22s
Playwright Tests / test-playwright (1, 2) (push) Failing after 2m42s
Playwright Tests / test-playwright (2, 2) (push) Failing after 1m41s
pre-commit / pre-commit (push) Failing after 2m53s
Test Backend / test-backend (push) Successful in 2m21s
Compose Smoke Test / test-compose (push) Successful in 31s
Playwright Tests / merge-reports (push) Failing after 1m6s
A cache hit still replays no emissions — those values were the story of an
execution that is not happening — but the run they were recorded in is now
written on the row (`run_node.cached_from`), and the metrics endpoints read the
series back from there. So a reused run answers `run.metrics("train.loss")`
with the same points the run that trained did, rather than looking like a run
that produced no numbers at all. Pointed at rather than copied: a sweep of 500
reusing one frozen node would otherwise duplicate its curve 500 times.
That needed the cross-flow restore fixed first. The cache key has no flow in
it while the stored outputs are named for the flow that produced them, so
`quick.prepare` getting a hit from `train` wrote `train.dataset` into `quick`'s
state and the next node was called without its argument. One rule now covers
both halves: `requalify` reads a name owned by one flow as the same name in
another, applied to the restored outputs, to the node id behind the pointer,
and to the series names on the way out. Reuse across flows is kept.
Also: `@run:<id>.<output>` and a bare `sha256:` digest resolve on every input,
not only artifacts. Chaining a run's json config into the next one from a shell
meant pasting the whole object inline, and the CLI could not even send the
spelling — `_coerce` died in `json.loads` before the engine saw it. Both
spellings are reserved on every input now, `str` included, and `_from_run`
returns whatever the run's result holds rather than only a reference.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dp9L6gakMVro1K2C5zdtBE
923 lines
36 KiB
Python
923 lines
36 KiB
Python
"""Runs: a flow taken from its inputs to its outputs, once.
|
|
|
|
A cascade is what an always-on flow does when a value arrives — it has no
|
|
beginning and no end worth recording. A *run* is the other shape the same
|
|
engine can take: parameters go in, the graph executes until it drains, and
|
|
what it produced is kept. That is what an ML experiment is, and what a
|
|
CI-style job is, so both are this one entity.
|
|
|
|
Three things make a run different from a cascade, and each is deliberate:
|
|
|
|
* **Its own state.** Message names are global keys, so two runs of one flow
|
|
sharing the engine's state would overwrite each other's values. A run gets a
|
|
state backend namespaced to itself, which is a constructor argument rather
|
|
than a change to the pipeline — every key the engine keeps, versions and
|
|
node memory included, already goes through that backend.
|
|
* **Its own record.** The event bus drops what it cannot keep up with, which
|
|
is right for a live canvas and wrong for a result. The driver writes the
|
|
run's rows itself, from the thread that is running it.
|
|
* **Its own durability.** The queue wakes an engine up; from the moment a run
|
|
starts, its database row is the truth. Redelivering hours of training
|
|
because an acknowledgement was late is not recovery, so a run is
|
|
acknowledged as soon as it is claimed and a stale lease — not an unacked
|
|
stream entry — is what marks a run its engine died in the middle of.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import copy
|
|
import hashlib
|
|
import json
|
|
import logging
|
|
import math
|
|
import os
|
|
import socket
|
|
import threading
|
|
import time
|
|
import uuid
|
|
from collections.abc import Callable
|
|
from concurrent.futures import ThreadPoolExecutor
|
|
from datetime import UTC, datetime, timedelta
|
|
from typing import Any
|
|
|
|
from sqlalchemy import update
|
|
from sqlalchemy.dialects.sqlite import insert as upsert
|
|
from sqlmodel import Session, col, select
|
|
|
|
from fluksio.core.db import engine as db_engine
|
|
from fluksio.flow.artifacts import ArtifactStore, is_reference, valid_digest
|
|
from fluksio.flow.controller import FlowController, RunContext
|
|
from fluksio.flow.messages import qualify
|
|
from fluksio.flow.pipeline import CacheHit, NodeOutcome, Pipeline
|
|
from fluksio.flow.queue import WorkItem, WorkQueue
|
|
from fluksio.flow.schemas import FlowDef
|
|
from fluksio.flow.state import MemoryState, StateBackend
|
|
from fluksio.models import Run, RunArtifact, RunMetric, RunNode
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
#: How long a finished run's state is kept before Redis drops it. Long enough
|
|
#: to look at what a failed run left behind, short enough not to accumulate.
|
|
RUN_STATE_TTL = 24 * 3600
|
|
#: How often a running run says it is still alive.
|
|
LEASE_INTERVAL_S = 20.0
|
|
#: A lease older than this belongs to an engine that is not coming back.
|
|
LEASE_STALE_S = 90.0
|
|
#: How often stale leases are looked for.
|
|
SWEEP_INTERVAL_S = 30.0
|
|
#: Runs driven at once. Node bodies are bounded by the worker pool anyway;
|
|
#: this only bounds how many graphs are in flight.
|
|
MAX_PARALLEL = 4
|
|
CLAIM_COUNT = 4
|
|
CLAIM_BLOCK_MS = 1000
|
|
ERROR_CAP = 2000
|
|
LOG_CAP = 8000
|
|
#: How much of a node's returned outputs is kept so a later run can restore
|
|
#: them. Past it the node simply is not cacheable — a cache is not a store,
|
|
#: and an artifact is how a large result is meant to travel.
|
|
OUTPUT_CAP = 256_000
|
|
#: Reported numbers held before they are written. A training loop reporting
|
|
#: every step must not be a round trip every step.
|
|
METRIC_BATCH = 500
|
|
METRIC_FLUSH_S = 2.0
|
|
#: How often a run waiting for a worker looks again.
|
|
WAIT_RETRY_S = 15.0
|
|
|
|
#: Where a run's state lives, so it can never collide with the engine's own.
|
|
RUN_NAMESPACE = "run"
|
|
|
|
|
|
class RunRejected(ValueError):
|
|
"""The run cannot be made: bad parameters, or a flow that cannot batch."""
|
|
|
|
|
|
def new_run_id() -> str:
|
|
"""Time-ordered, so the newest runs sort last without reading a column."""
|
|
return f"{int(time.time() * 1000):013d}-{uuid.uuid4().hex[:8]}"
|
|
|
|
|
|
def digest_of(params: dict[str, Any], seed: int | None) -> str:
|
|
"""What identifies a run's inputs: same parameters, same digest."""
|
|
canonical = json.dumps(
|
|
{"params": params, "seed": seed}, sort_keys=True, separators=(",", ":")
|
|
)
|
|
return hashlib.sha256(canonical.encode()).hexdigest()
|
|
|
|
|
|
def batch_issues(flow: FlowDef) -> list[str]:
|
|
"""Why this flow cannot be run as a batch, if it cannot.
|
|
|
|
One thing genuinely breaks: a port with a discretization interval holds
|
|
values back for a timer to release, and a run has no timer — the engine
|
|
would drop them instead of delaying them. On a *streaming* port that is
|
|
exactly right and is how you keep a chart from being flooded: the run's
|
|
history keeps every value, and the interval only thins what is published.
|
|
Anywhere else it is a message quietly going missing.
|
|
|
|
A delay node is fine; without a queue to defer into it simply sleeps,
|
|
which in a run is what was asked for.
|
|
"""
|
|
issues: list[str] = []
|
|
for node in flow.nodes:
|
|
for spec in list(node.requires) + list(node.provides):
|
|
if spec.interval > 0 and not spec.stream:
|
|
issues.append(
|
|
f"Node '{node.id}' rate-limits '{spec.port or spec.name}'. "
|
|
"A run has no timer to release what that holds back, so "
|
|
"the value would be dropped. Remove the interval, or mark "
|
|
"the port as streaming if it is a curve being thinned out."
|
|
)
|
|
return issues
|
|
|
|
|
|
def required_labels(flow: FlowDef) -> list[str]:
|
|
"""Worker labels this flow cannot run without.
|
|
|
|
Only the nodes that *require* their device. One that merely prefers it runs
|
|
on the engine when no such worker is attached, so holding the whole run
|
|
back to wait for one would be waiting for something it does not need.
|
|
"""
|
|
return sorted(
|
|
{
|
|
node.device
|
|
for node in flow.nodes
|
|
if node.device and node.device.strip() and node.device_policy == "require"
|
|
}
|
|
)
|
|
|
|
|
|
#: What a run's output is called from outside it: ``@run:<id>.<output>``.
|
|
RUN_REF_PREFIX = "@run:"
|
|
|
|
|
|
def resolve_references(
|
|
flow: FlowDef, params: dict[str, Any], artifacts: ArtifactStore | None = None
|
|
) -> dict[str, Any]:
|
|
"""Turn the text spelling of a run's output into the value itself.
|
|
|
|
A python caller hands one run's output straight to the next, because it has
|
|
the value in its hand. A shell does not, and pasting a config object or an
|
|
artifact reference is not a command anyone wants to type — so any declared
|
|
input also takes ``@run:<id>.<output>``, naming what a run produced, and a
|
|
bare ``sha256:...`` digest naming bytes. Resolved here rather than in each
|
|
client, so the CLI, the browser and a python caller all mean the same thing
|
|
by the same string.
|
|
|
|
Both spellings are therefore reserved on every input, ``str`` included: an
|
|
input that has to carry one of them literally is asking for a value this
|
|
engine reads as a name.
|
|
"""
|
|
declared = {one.spec.name for one in flow.inputs}
|
|
pending = {
|
|
key: value
|
|
for key, value in params.items()
|
|
if key in declared
|
|
and isinstance(value, str)
|
|
and (value.startswith(RUN_REF_PREFIX) or valid_digest(value))
|
|
}
|
|
if not pending:
|
|
return params
|
|
|
|
resolved = dict(params)
|
|
with Session(db_engine) as session:
|
|
for key, text in pending.items():
|
|
if text.startswith(RUN_REF_PREFIX):
|
|
resolved[key] = _from_run(session, key, text[len(RUN_REF_PREFIX) :])
|
|
else:
|
|
resolved[key] = _from_digest(session, key, text, artifacts)
|
|
return resolved
|
|
|
|
|
|
def _from_run(session: Session, key: str, spelling: str) -> Any:
|
|
"""``<run id>.<output>`` as the value that run produced."""
|
|
run_id, _, output = spelling.partition(".")
|
|
if not run_id or not output:
|
|
raise RunRejected(
|
|
f"Parameter '{key}': '{RUN_REF_PREFIX}{spelling}' names no output — "
|
|
f"write {RUN_REF_PREFIX}<run id>.<output>"
|
|
)
|
|
run = session.get(Run, run_id)
|
|
if run is None:
|
|
raise RunRejected(f"Parameter '{key}': there is no run '{run_id}'")
|
|
|
|
# The run's own result first: that is the value as its producer made it,
|
|
# an artifact's file name and all. Copied, because what comes back is
|
|
# handed on as this run's parameter and must not alias the other run's row.
|
|
result = run.result or {}
|
|
if output in result:
|
|
return copy.deepcopy(result[output])
|
|
|
|
# The artifact rows are the fallback: bytes a node made that the flow never
|
|
# declared as an output. They carry the message name instead, which loads
|
|
# the same bytes either way.
|
|
rows = session.exec(
|
|
select(RunArtifact).where(col(RunArtifact.run_id) == run_id)
|
|
).all()
|
|
for row in rows:
|
|
if output in (row.name, row.name.rsplit(".", 1)[-1]):
|
|
return {
|
|
"digest": row.digest,
|
|
"size": row.size,
|
|
"media_type": row.media_type or "application/octet-stream",
|
|
"name": row.name,
|
|
}
|
|
known = ", ".join(sorted({*result, *(row.name for row in rows)})) or "none"
|
|
raise RunRejected(
|
|
f"Parameter '{key}': run '{run_id}' has no output '{output}' (it made: {known})"
|
|
)
|
|
|
|
|
|
def _from_digest(
|
|
session: Session, key: str, digest: str, artifacts: ArtifactStore | None
|
|
) -> dict[str, Any]:
|
|
"""A bare digest as a reference, with the size the store needs."""
|
|
row = session.exec(
|
|
select(RunArtifact)
|
|
.where(col(RunArtifact.digest) == digest)
|
|
.order_by(col(RunArtifact.run_id).desc())
|
|
).first()
|
|
if row is None:
|
|
raise RunRejected(
|
|
f"Parameter '{key}': no run has produced '{digest}', so there is "
|
|
"nothing here under that digest"
|
|
)
|
|
if artifacts is not None and artifacts.path(digest) is None:
|
|
raise RunRejected(
|
|
f"Parameter '{key}': '{digest}' is known but its bytes are gone "
|
|
"from this installation's store"
|
|
)
|
|
return {
|
|
"digest": row.digest,
|
|
"size": row.size,
|
|
"media_type": row.media_type or "application/octet-stream",
|
|
"name": row.name,
|
|
}
|
|
|
|
|
|
def seed_values(
|
|
flow: FlowDef, params: dict[str, Any], seed: int | None = None
|
|
) -> dict[str, Any]:
|
|
"""Turn a run's parameters into the messages the flow starts from.
|
|
|
|
A run's ``seed`` fills an input of that name when the flow declares one, so
|
|
the field that distinguishes two runs of one configuration is the same
|
|
number the flow draws from — otherwise it would only look like it was, and
|
|
a sweep over seeds would run the same experiment N times. An explicit
|
|
parameter still wins, and a flow that declares no ``seed`` ignores it.
|
|
"""
|
|
specs = {declared.spec.name: declared.spec for declared in flow.inputs}
|
|
if seed is not None and "seed" in specs and "seed" not in params:
|
|
params = {**params, "seed": seed}
|
|
values: dict[str, Any] = {}
|
|
for key, value in params.items():
|
|
spec = specs.get(key)
|
|
if spec is None:
|
|
known = ", ".join(sorted(specs)) or "none"
|
|
raise RunRejected(
|
|
f"'{key}' is not an input of flow '{flow.name}' (it declares: {known})"
|
|
)
|
|
try:
|
|
spec.check(value)
|
|
except TypeError as exc:
|
|
raise RunRejected(f"Parameter '{key}': {exc}") from exc
|
|
values[qualify(flow.name, spec.name)] = value
|
|
return values
|
|
|
|
|
|
def collect_result(flow: FlowDef, state: StateBackend) -> dict[str, Any]:
|
|
"""What the run produced, keyed by message name without the flow prefix."""
|
|
prefix = f"{flow.name}."
|
|
if flow.outputs:
|
|
names = [qualify(flow.name, name) for name in flow.outputs]
|
|
else:
|
|
# Everything the flow ended up holding. The engine's own bookkeeping is
|
|
# keyed by `__thing__:message`, so it never starts with the flow name.
|
|
names = sorted(key for key in state.keys() if key.startswith(prefix))
|
|
result: dict[str, Any] = {}
|
|
for name in names:
|
|
if name in state:
|
|
result[name[len(prefix) :] if name.startswith(prefix) else name] = state[
|
|
name
|
|
]
|
|
return result
|
|
|
|
|
|
class MetricSink:
|
|
"""Keeps the series a run's streaming outputs traced out.
|
|
|
|
A run's metrics are not logged; they are the numbers its nodes published on
|
|
the way to finishing. This watches the emissions, keeps the numeric ones,
|
|
and writes them in batches — synchronously rather than over the event bus,
|
|
which drops what it cannot keep up with, and a training curve with holes in
|
|
it is not a result.
|
|
|
|
The step is the count of emissions on that message. A node that publishes
|
|
every tenth training step therefore has steps 0, 1, 2 rather than 0, 10,
|
|
20 — a faithful x-axis of its own emissions, not of the loop inside it.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
run_id: str,
|
|
batch: int = METRIC_BATCH,
|
|
interval: float = METRIC_FLUSH_S,
|
|
) -> None:
|
|
self.run_id = run_id
|
|
self._batch = batch
|
|
self._interval = interval
|
|
self._rows: dict[tuple[str, int], RunMetric] = {}
|
|
self._steps: dict[str, int] = {}
|
|
self._last_flush = time.monotonic()
|
|
self._lock = threading.Lock()
|
|
|
|
def handle(self, node_id: str, outputs: dict[str, Any]) -> None:
|
|
"""One emission: every number in it belongs to this run's history."""
|
|
now = time.time()
|
|
rows: list[RunMetric] = []
|
|
with self._lock:
|
|
for name, value in outputs.items():
|
|
if not isinstance(value, (int, float)) or isinstance(value, bool):
|
|
# A checkpoint or a record is on the run some other way —
|
|
# as an artifact, or as its result. Only numbers are series.
|
|
continue
|
|
if not math.isfinite(value):
|
|
# The ports refuse these, so one here came from somewhere
|
|
# that does not go through them. A column that cannot hold
|
|
# it would take the whole batch down with it.
|
|
continue
|
|
step = self._steps.get(name, -1) + 1
|
|
self._steps[name] = step
|
|
row = RunMetric(
|
|
run_id=self.run_id,
|
|
name=name[:128],
|
|
step=step,
|
|
node=node_id[:255],
|
|
ts=now,
|
|
value=float(value),
|
|
)
|
|
self._rows[(row.name, row.step)] = row
|
|
due = (
|
|
len(self._rows) >= self._batch
|
|
or time.monotonic() - self._last_flush >= self._interval
|
|
)
|
|
if due:
|
|
rows = list(self._rows.values())
|
|
self._rows.clear()
|
|
self._last_flush = time.monotonic()
|
|
if rows:
|
|
self._write(rows)
|
|
|
|
def flush(self) -> None:
|
|
with self._lock:
|
|
rows = list(self._rows.values())
|
|
self._rows.clear()
|
|
self._last_flush = time.monotonic()
|
|
if rows:
|
|
self._write(rows)
|
|
|
|
def _write(self, rows: list[RunMetric]) -> None:
|
|
try:
|
|
with Session(db_engine) as session:
|
|
statement = upsert(RunMetric).values([row.model_dump() for row in rows])
|
|
session.exec(
|
|
statement.on_conflict_do_update(
|
|
index_elements=["run_id", "name", "step"],
|
|
set_={
|
|
"value": statement.excluded.value,
|
|
"ts": statement.excluded.ts,
|
|
"node": statement.excluded.node,
|
|
},
|
|
)
|
|
)
|
|
session.commit()
|
|
except Exception:
|
|
logger.exception(
|
|
"Could not write %d metric(s) of %s", len(rows), self.run_id
|
|
)
|
|
|
|
|
|
def origin_commit(flow: FlowDef) -> str:
|
|
"""The user repository's commit, for a flow declared in code elsewhere.
|
|
|
|
Marked dirty when the tree had uncommitted changes at sync, because then
|
|
the hash names something other than what actually ran.
|
|
"""
|
|
origin = flow.origin
|
|
if origin is None or not origin.commit:
|
|
return ""
|
|
return f"{origin.commit}-dirty" if origin.dirty else origin.commit
|
|
|
|
|
|
def _cacheable(outcome: NodeOutcome) -> str | None:
|
|
"""A node's outputs as stored, or None when it may not be reused."""
|
|
if not outcome.ok or not outcome.cache_key:
|
|
return None
|
|
try:
|
|
text = json.dumps(outcome.output_values, sort_keys=True, separators=(",", ":"))
|
|
except (TypeError, ValueError):
|
|
return None
|
|
return text if len(text) <= OUTPUT_CAP else None
|
|
|
|
|
|
class RunCache:
|
|
"""What earlier runs already worked out, looked up by cache key.
|
|
|
|
The rows are the run history itself — a node that ran successfully and
|
|
whose outputs were small enough to keep is a cache entry, without a store
|
|
of its own. An entry whose artifacts have gone from the content store is
|
|
not a hit: the reference would name bytes nobody can open.
|
|
"""
|
|
|
|
def __init__(self, artifacts: ArtifactStore | None = None) -> None:
|
|
self.artifacts = artifacts
|
|
|
|
def lookup(self, key: str) -> CacheHit | None:
|
|
if not key:
|
|
return None
|
|
try:
|
|
with Session(db_engine) as session:
|
|
rows = session.exec(
|
|
select(RunNode)
|
|
.where(
|
|
col(RunNode.cache_key) == key,
|
|
col(RunNode.status).in_(("ok", "cached")),
|
|
col(RunNode.outputs).is_not(None),
|
|
)
|
|
# Run ids are time-ordered, so this is the most recent
|
|
# first. A handful is enough to get past entries whose
|
|
# artifacts have been collected.
|
|
.order_by(col(RunNode.run_id).desc())
|
|
.limit(5)
|
|
).all()
|
|
for row in rows:
|
|
outputs = json.loads(row.outputs or "null")
|
|
if outputs and self._unrestorable(outputs):
|
|
continue
|
|
source = session.get(Run, row.run_id)
|
|
return CacheHit(
|
|
flow=source.flow if source is not None else "",
|
|
outputs=outputs,
|
|
# A row that was itself restored holds no series, so
|
|
# follow its pointer rather than adding a hop to it.
|
|
metrics_run=row.cached_from or row.run_id,
|
|
)
|
|
except Exception:
|
|
# A cache that cannot answer is a cache miss, never a failed run.
|
|
logger.exception("Cache lookup failed; running the node instead")
|
|
return None
|
|
|
|
def _unrestorable(self, outputs: dict[str, Any]) -> bool:
|
|
return any(
|
|
is_reference(value)
|
|
and (
|
|
self.artifacts is None
|
|
or self.artifacts.path(str(value["digest"])) is None
|
|
)
|
|
for value in outputs.values()
|
|
)
|
|
|
|
|
|
class RunService:
|
|
"""Accepts runs, drives them, and writes down what they did."""
|
|
|
|
def __init__(
|
|
self,
|
|
controller: FlowController,
|
|
queue: WorkQueue,
|
|
state_factory: Callable[[str], StateBackend] | None = None,
|
|
parallel: int = MAX_PARALLEL,
|
|
artifacts: ArtifactStore | None = None,
|
|
) -> None:
|
|
self.controller = controller
|
|
self.queue = queue
|
|
self._artifacts = artifacts
|
|
self._cache = RunCache(artifacts)
|
|
# Without one, a run gets a private in-memory state — which is exactly
|
|
# the isolation it wants, minus surviving the process.
|
|
self._state_factory = state_factory or (lambda _ns: MemoryState())
|
|
self.engine_name = f"{socket.gethostname()}-{os.getpid()}"[:64]
|
|
self._pool = ThreadPoolExecutor(max_workers=parallel, thread_name_prefix="run")
|
|
self._stop = threading.Event()
|
|
self._consumer: threading.Thread | None = None
|
|
self._keeper: threading.Thread | None = None
|
|
# Runs this process is driving, and the pipeline each is running, so a
|
|
# cancel has something to hold on to.
|
|
self._active: dict[str, Pipeline] = {}
|
|
self._cancelled: set[str] = set()
|
|
self._lock = threading.Lock()
|
|
|
|
# -------------------------------------------------------------------------
|
|
# Lifecycle
|
|
# -------------------------------------------------------------------------
|
|
|
|
def start(self) -> None:
|
|
if self._consumer is not None:
|
|
return
|
|
self._consumer = threading.Thread(
|
|
target=self._consume, name="run-consumer", daemon=True
|
|
)
|
|
self._consumer.start()
|
|
self._keeper = threading.Thread(
|
|
target=self._keep_leases, name="run-leases", daemon=True
|
|
)
|
|
self._keeper.start()
|
|
|
|
def stop(self) -> None:
|
|
self._stop.set()
|
|
for thread in (self._consumer, self._keeper):
|
|
if thread is not None:
|
|
thread.join(timeout=5)
|
|
self._consumer = None
|
|
self._keeper = None
|
|
self._pool.shutdown(wait=False)
|
|
self.queue.close()
|
|
|
|
def alive(self) -> bool:
|
|
return self._consumer is not None and self._consumer.is_alive()
|
|
|
|
# -------------------------------------------------------------------------
|
|
# Accepting work
|
|
# -------------------------------------------------------------------------
|
|
|
|
def submit(
|
|
self,
|
|
flow_name: str,
|
|
params: dict[str, Any] | None = None,
|
|
seed: int | None = None,
|
|
group_id: str | None = None,
|
|
cause: str = "api",
|
|
actor: str = "",
|
|
draft: bool = False,
|
|
no_cache: bool = False,
|
|
) -> Run:
|
|
"""Journal a run and wake an engine up for it. Never blocks on it."""
|
|
flow = self.controller.store.read_flow(flow_name, draft=draft)
|
|
issues = batch_issues(flow)
|
|
if issues:
|
|
raise RunRejected(" ".join(issues))
|
|
params = params or {}
|
|
# Resolved before it is stored, so what the run records is the same
|
|
# reference a python caller would have passed and every later reader —
|
|
# the digest, the cache, the run detail — sees one spelling.
|
|
params = resolve_references(flow, params, self._artifacts)
|
|
# Checked here rather than in the driver: a caller who mistyped a
|
|
# parameter should be told now, not by a run that fails in a minute.
|
|
seed_values(flow, params, seed)
|
|
|
|
run = Run(
|
|
id=new_run_id(),
|
|
flow=flow.name,
|
|
flow_version=flow.version,
|
|
commit=self.controller.store.head(),
|
|
origin_commit=origin_commit(flow),
|
|
params=params,
|
|
params_digest=digest_of(params, seed),
|
|
seed=seed,
|
|
group_id=group_id,
|
|
cause=cause,
|
|
no_cache=no_cache,
|
|
status="queued",
|
|
labels=required_labels(flow),
|
|
created_at=datetime.now(UTC),
|
|
actor=actor,
|
|
)
|
|
with Session(db_engine) as session:
|
|
session.add(run)
|
|
session.commit()
|
|
session.refresh(run)
|
|
|
|
self.queue.add(WorkItem(kind="run", node="", flow=flow.name, run_id=run.id))
|
|
return run
|
|
|
|
def cancel(self, run_id: str) -> bool:
|
|
"""Stop a run: kill what it is executing, schedule nothing further."""
|
|
with self._lock:
|
|
pipeline = self._active.get(run_id)
|
|
if pipeline is None:
|
|
# Not running here — if it is still queued, refusing to start
|
|
# is all the cancelling it needs.
|
|
cancelled = self._finish_queued(run_id)
|
|
if cancelled:
|
|
self._cancelled.add(run_id)
|
|
return cancelled
|
|
self._cancelled.add(run_id)
|
|
# The gate first, so nothing new is submitted while the running nodes
|
|
# are being killed; a gated node is never handed to the executor, so
|
|
# the graph drains instead of going further.
|
|
pipeline.pause(self._flow_of(run_id) or "")
|
|
workers = self.controller.workers
|
|
if workers is not None:
|
|
# Keyed by run, so a sweep cancelling one config leaves the others
|
|
# training.
|
|
workers.cancel_run(run_id)
|
|
if self.controller.remote is not None:
|
|
self.controller.remote.cancel_run(run_id)
|
|
return True
|
|
|
|
def _flow_of(self, run_id: str) -> str | None:
|
|
with Session(db_engine) as session:
|
|
run = session.get(Run, run_id)
|
|
return run.flow if run else None
|
|
|
|
def _finish_queued(self, run_id: str) -> bool:
|
|
with Session(db_engine) as session:
|
|
result = session.exec(
|
|
update(Run)
|
|
.where(col(Run.id) == run_id, col(Run.status) == "queued")
|
|
.values(
|
|
status="cancelled",
|
|
finished_at=datetime.now(UTC),
|
|
status_reason="Cancelled before it started",
|
|
)
|
|
)
|
|
session.commit()
|
|
return bool(result.rowcount)
|
|
|
|
# -------------------------------------------------------------------------
|
|
# Threads
|
|
# -------------------------------------------------------------------------
|
|
|
|
def _consume(self) -> None:
|
|
failures = 0
|
|
while not self._stop.is_set():
|
|
try:
|
|
# Runs put back to wait for a worker come due here. The claim
|
|
# below blocks for a second, so this is about once a second.
|
|
self.queue.move_due(time.time())
|
|
items = self.queue.claim(CLAIM_COUNT, CLAIM_BLOCK_MS)
|
|
failures = 0
|
|
except Exception as exc:
|
|
failures += 1
|
|
logger.error("Could not claim runs: %s", exc)
|
|
self._stop.wait(min(30.0, 2.0**failures))
|
|
continue
|
|
for item in items:
|
|
if not item.run_id:
|
|
self.queue.ack(item)
|
|
continue
|
|
missing = self._missing_labels(item.run_id)
|
|
if missing:
|
|
# Left in the queue rather than failed: submitting a run
|
|
# before turning the GPU box on is a normal way to work, and
|
|
# the run says what it is waiting for while it waits.
|
|
self._waiting(item.run_id, missing)
|
|
self._defer(item)
|
|
continue
|
|
# Acknowledged before it runs: from here on the row is the
|
|
# record, and a lease that stops moving is what says otherwise.
|
|
self.queue.ack(item)
|
|
try:
|
|
self._pool.submit(self._drive, item.run_id)
|
|
except RuntimeError:
|
|
logger.warning("Run %s not started: shutting down", item.run_id)
|
|
|
|
def _missing_labels(self, run_id: str) -> list[str]:
|
|
"""Worker labels this run needs that nothing attached carries."""
|
|
with Session(db_engine) as session:
|
|
run = session.get(Run, run_id)
|
|
needed = list(run.labels) if run else []
|
|
if not needed:
|
|
return []
|
|
hub = self.controller.remote
|
|
available = hub.labels() | {w.name for w in hub.workers()} if hub else set()
|
|
# A node that only prefers its label runs locally instead, so it is not
|
|
# a reason to hold the run back; that is decided per node at call time.
|
|
return sorted(set(needed) - available)
|
|
|
|
def _waiting(self, run_id: str, missing: list[str]) -> None:
|
|
reason = f"Waiting for a worker labelled {', '.join(missing)}"
|
|
try:
|
|
with Session(db_engine) as session:
|
|
session.exec(
|
|
update(Run)
|
|
.where(col(Run.id) == run_id, col(Run.status) == "queued")
|
|
.values(status_reason=reason)
|
|
)
|
|
session.commit()
|
|
except Exception:
|
|
logger.exception("Could not record what run %s is waiting for", run_id)
|
|
|
|
def _defer(self, item: WorkItem) -> None:
|
|
"""Put an item back for later, and let go of this delivery."""
|
|
try:
|
|
self.queue.add_delayed(item, time.time() + WAIT_RETRY_S)
|
|
self.queue.ack(item)
|
|
except Exception:
|
|
logger.exception("Could not defer run %s", item.run_id)
|
|
|
|
def _keep_leases(self) -> None:
|
|
"""Say the local runs are alive, and clean up after engines that died."""
|
|
last_sweep = 0.0
|
|
while not self._stop.is_set():
|
|
self._stop.wait(LEASE_INTERVAL_S)
|
|
if self._stop.is_set():
|
|
break
|
|
with self._lock:
|
|
mine = list(self._active)
|
|
now = datetime.now(UTC)
|
|
try:
|
|
if mine:
|
|
with Session(db_engine) as session:
|
|
session.exec(
|
|
update(Run)
|
|
.where(col(Run.id).in_(mine))
|
|
.values(lease_at=now)
|
|
)
|
|
session.commit()
|
|
if time.monotonic() - last_sweep >= SWEEP_INTERVAL_S:
|
|
last_sweep = time.monotonic()
|
|
self._sweep_abandoned(now)
|
|
except Exception:
|
|
logger.exception("Could not refresh run leases")
|
|
|
|
def _sweep_abandoned(self, now: datetime) -> None:
|
|
cutoff = now - timedelta(seconds=LEASE_STALE_S)
|
|
with Session(db_engine) as session:
|
|
result = session.exec(
|
|
update(Run)
|
|
.where(
|
|
col(Run.status) == "running",
|
|
col(Run.lease_at) < cutoff,
|
|
)
|
|
.values(
|
|
status="abandoned",
|
|
finished_at=now,
|
|
status_reason="The engine running it stopped reporting",
|
|
)
|
|
)
|
|
session.commit()
|
|
if result.rowcount:
|
|
logger.warning("Marked %d run(s) abandoned", result.rowcount)
|
|
|
|
# -------------------------------------------------------------------------
|
|
# Driving one run
|
|
# -------------------------------------------------------------------------
|
|
|
|
def _claim(self, run_id: str) -> Run | None:
|
|
"""Take the run, or leave it: whoever moves it out of `queued` owns it.
|
|
|
|
The compare-and-swap is what makes a redelivered item harmless — the
|
|
second engine to arrive updates nothing and walks away.
|
|
"""
|
|
now = datetime.now(UTC)
|
|
with Session(db_engine) as session:
|
|
result = session.exec(
|
|
update(Run)
|
|
.where(col(Run.id) == run_id, col(Run.status) == "queued")
|
|
.values(
|
|
status="running",
|
|
started_at=now,
|
|
lease_at=now,
|
|
engine=self.engine_name,
|
|
)
|
|
)
|
|
session.commit()
|
|
if not result.rowcount:
|
|
return None
|
|
return session.exec(select(Run).where(col(Run.id) == run_id)).first()
|
|
|
|
def _drive(self, run_id: str) -> None:
|
|
run = self._claim(run_id)
|
|
if run is None:
|
|
return
|
|
started = time.perf_counter()
|
|
status = "ok"
|
|
reason = ""
|
|
result: dict[str, Any] = {}
|
|
state: StateBackend | None = None
|
|
errors = 0
|
|
|
|
def observe(outcome: NodeOutcome) -> None:
|
|
nonlocal errors
|
|
if not outcome.ok:
|
|
errors += 1
|
|
self._record_node(run_id, outcome)
|
|
|
|
sink = MetricSink(run_id)
|
|
try:
|
|
flow = self.controller.store.read_flow(run.flow)
|
|
state = self._state_factory(f"{RUN_NAMESPACE}:{run_id}")
|
|
pipeline = self.controller.build_run_pipeline(
|
|
flow,
|
|
state=state,
|
|
observer=observe,
|
|
emission_observer=sink.handle,
|
|
run=RunContext(run_id=run_id),
|
|
run_cache=None if run.no_cache else self._cache,
|
|
)
|
|
with self._lock:
|
|
self._active[run_id] = pipeline
|
|
if run_id in self._cancelled:
|
|
pipeline.pause(flow.name)
|
|
self._publish(run, "run_started")
|
|
pipeline.run(seed_values(flow, run.params, run.seed))
|
|
result = collect_result(flow, state)
|
|
with self._lock:
|
|
cancelled = run_id in self._cancelled
|
|
if cancelled:
|
|
status = "cancelled"
|
|
reason = "Cancelled while running"
|
|
elif errors:
|
|
status = "error"
|
|
reason = f"{errors} node(s) failed"
|
|
except Exception as exc:
|
|
logger.exception("Run %s failed", run_id)
|
|
status = "error"
|
|
reason = f"{type(exc).__name__}: {exc}"[:1024]
|
|
finally:
|
|
with self._lock:
|
|
self._active.pop(run_id, None)
|
|
self._cancelled.discard(run_id)
|
|
# Whatever the last batch was holding belongs to this run's record.
|
|
sink.flush()
|
|
duration = round((time.perf_counter() - started) * 1000, 2)
|
|
self._finish(run_id, status, reason, result, duration)
|
|
run.status = status
|
|
self._publish(run, "run_finished")
|
|
# Its values were only ever this run's; nothing reads them once it
|
|
# has a result. On Redis the namespace would expire anyway.
|
|
if state is not None and status != "error":
|
|
try:
|
|
state.clear()
|
|
except Exception:
|
|
logger.warning("Could not clear state of run %s", run_id)
|
|
|
|
def _record_node(self, run_id: str, outcome: NodeOutcome) -> None:
|
|
outputs = _cacheable(outcome)
|
|
row = RunNode(
|
|
run_id=run_id,
|
|
node=outcome.node[:255],
|
|
status="cached" if outcome.cached else ("ok" if outcome.ok else "error"),
|
|
started_at=datetime.now(UTC),
|
|
duration_ms=outcome.duration_ms,
|
|
error=outcome.error[:ERROR_CAP],
|
|
logs=outcome.logs[:LOG_CAP],
|
|
cached_from=outcome.cached_from,
|
|
# Together or not at all: a row carrying a key must be one a
|
|
# lookup can actually restore from.
|
|
cache_key=outcome.cache_key if outputs is not None else "",
|
|
outputs=outputs,
|
|
)
|
|
try:
|
|
with Session(db_engine) as session:
|
|
session.merge(row)
|
|
for message, ref in outcome.artifacts.items():
|
|
session.merge(
|
|
RunArtifact(
|
|
run_id=run_id,
|
|
name=message[:255],
|
|
node=outcome.node[:255],
|
|
digest=str(ref.get("digest") or "")[:71],
|
|
size=int(ref.get("size") or 0),
|
|
media_type=str(
|
|
ref.get("media_type") or "application/octet-stream"
|
|
)[:128],
|
|
)
|
|
)
|
|
session.commit()
|
|
except Exception:
|
|
logger.exception(
|
|
"Could not record node '%s' of run %s", outcome.node, run_id
|
|
)
|
|
|
|
def _finish(
|
|
self,
|
|
run_id: str,
|
|
status: str,
|
|
reason: str,
|
|
result: dict[str, Any],
|
|
duration_ms: float,
|
|
) -> None:
|
|
try:
|
|
with Session(db_engine) as session:
|
|
session.exec(
|
|
update(Run)
|
|
.where(col(Run.id) == run_id)
|
|
.values(
|
|
status=status,
|
|
status_reason=reason,
|
|
result=result,
|
|
duration_ms=duration_ms,
|
|
finished_at=datetime.now(UTC),
|
|
)
|
|
)
|
|
session.commit()
|
|
except Exception:
|
|
logger.exception("Could not close run %s", run_id)
|
|
|
|
def _publish(self, run: Run, kind: str) -> None:
|
|
self._publish_event(
|
|
{
|
|
"type": kind,
|
|
"flow": run.flow,
|
|
"run": run.id,
|
|
"status": run.status,
|
|
"group": run.group_id or "",
|
|
"ts": time.time(),
|
|
}
|
|
)
|
|
|
|
def _publish_event(self, event: dict[str, Any]) -> None:
|
|
if self.controller.events is not None:
|
|
self.controller.events.publish(event)
|