Compare commits

...
Author SHA1 Message Date
Hongsheng Liu bfe1df6bce Merge branch 'main' into codex-hwr-domain-timeout 2026-09-10 13:29:45 +08:00
Hongsheng Liu d455411e9e [Bugfix] Bound HWR domain initialization lock waits
Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>
2026-09-06 16:06:21 +08:00
5 changed files with 134 additions and 9 deletions

No files matched your search

+9 -2
View File
@@ -313,8 +313,15 @@ One process per exact identity owns a build; other workers wait and then acquire
leases for the published artifact. Publication is invisible until all payloads
and metadata are validated, hashed, fsynced, and atomically renamed.
`coordination_timeout_seconds` bounds filesystem lock acquisition. It does not
cancel synchronous validation, a producer that has already started, or atomic
`coordination_timeout_seconds` bounds domain-initialization and lookup/build
lock acquisition. Store construction and each later resolution or publication
operation have separate budgets from the same wait policy, rather than one
end-to-end startup deadline. A domain-init timeout follows the retryable domain
failure policy below. After contention ends, a fresh runtime construction can
retry initialization; the timed-out runtime retains its original failure.
The coordination budget does not cancel filesystem I/O, synchronous validation,
a producer that has already started, or atomic
publication. A hung in-process producer therefore blocks its owning process and
must be handled by external process supervision. Enforceable producer
cancellation requires a future process-isolated producer contract.
+10 -2
View File
@@ -136,8 +136,16 @@ so only when it is marked retryable; unsupported capabilities, identity
collisions, producer failures, and publication failures remain visible as
startup failures.
`coordination_timeout_seconds` bounds acquisition of lookup and build locks; it
is not a hard wall-clock deadline for in-process validation, production, or
`coordination_timeout_seconds` bounds acquisition of the domain-initialization
lock and lookup/build locks. Store construction and each later resolution or
publication operation receive separate coordination budgets from the same
`WaitPolicy`; they do not share an end-to-end startup deadline. A domain-init
lock timeout is a retryable domain failure: preferred mode may fall back and
required mode fails. Retrying initialization requires constructing a fresh
runtime after contention ends; a waiter never cancels or unlocks the owner.
The coordination budget is not a hard wall-clock deadline for filesystem I/O,
in-process validation, production, or
atomic publication. Once this V1 implementation becomes the producer, the
synchronous producer and publication run to completion. A hung producer blocks
its owning process and requires external process supervision; enforceable
@@ -4,6 +4,10 @@
from __future__ import annotations
import multiprocessing as mp
import threading
import time
from multiprocessing.connection import Connection
from pathlib import Path
from typing import cast
@@ -15,6 +19,7 @@ from vllm_omni.host_weight_runtime import (
CanonicalJson,
ComponentIdentity,
FailureCode,
HostWeightError,
HostWeightFailure,
HostWeightLease,
HostWeightLeaseCarrier,
@@ -40,12 +45,14 @@ from vllm_omni.host_weight_runtime import (
StoreStatus,
TensorWriteSpec,
ValidationLevel,
WaitPolicy,
WeightArtifactIdentity,
WeightProductionSpec,
WeightRepresentation,
WeightSourceIdentity,
)
from vllm_omni.host_weight_runtime.filesystem import FilesystemHostWeightStore, detect_storage_class
from vllm_omni.host_weight_runtime.filesystem.locks import FileLock, lock_is_active
pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
@@ -518,6 +525,107 @@ def test_retryable_store_initialization_failure_obeys_runtime_mode(
assert resolution.report.attempts[0].failure.retryable
def _initialize_contended_runtime(config: HostWeightRuntimeConfig, output: Connection) -> None:
try:
output.send("starting")
started = time.monotonic()
runtime = HostWeightRuntime.from_config(config)
resolution = runtime.resolve(_identity())
try:
output.send((time.monotonic() - started, resolution.report))
finally:
if resolution.lease is not None:
resolution.lease.close()
finally:
output.close()
@pytest.mark.parametrize(
("mode", "expected"),
[
(RuntimeMode.PREFERRED, ResolutionOutcome.CANONICAL_FALLBACK),
(RuntimeMode.REQUIRED, ResolutionOutcome.FAILED),
],
)
def test_domain_lock_timeout_and_fresh_runtime_recovery(
tmp_path: Path, mode: RuntimeMode, expected: ResolutionOutcome
) -> None:
config = HostWeightRuntimeConfig(mode=mode, domain=_domain(tmp_path), wait=WaitPolicy(0.1))
initial = HostWeightRuntime.from_config(config)
built = initial.resolve(_identity(), producer=CountingProducer(_identity()))
assert built.lease is not None
built.lease.close()
metadata = {name: (tmp_path / name).read_bytes() for name in ("domain.json", "domain-policy.json")}
lock_path = tmp_path / "locks" / "domain-init.lock"
ctx = mp.get_context("spawn")
reader, writer = ctx.Pipe(duplex=False)
process = ctx.Process(target=_initialize_contended_runtime, args=(config, writer))
with FileLock(lock_path, exclusive=True, deadline=None):
process.start()
writer.close()
try:
# Import/bootstrap is outside the short coordination budget.
assert reader.poll(60), "waiter did not start"
assert reader.recv() == "starting"
assert reader.poll(5), "domain initialization did not respect the coordination timeout"
elapsed, report = reader.recv()
assert 0.1 <= elapsed < 5
assert report.outcome is expected
failure = report.attempts[0].failure
assert failure is not None
assert failure.stage is ResolutionStage.DOMAIN
assert failure.code is FailureCode.DOMAIN_UNAVAILABLE
assert failure.retryable
assert "timed out waiting for" in failure.message
assert "domain-init.lock" in failure.message
assert lock_is_active(lock_path), "waiter must not unlock the owner"
process.join(5)
assert process.exitcode == 0
finally:
if process.is_alive():
process.kill()
process.join(5)
process.close()
reader.close()
assert {name: (tmp_path / name).read_bytes() for name in metadata} == metadata
recovered = HostWeightRuntime.from_config(config).resolve(_identity())
assert recovered.report.outcome is ResolutionOutcome.LOCAL_HIT
assert recovered.lease is not None
recovered.lease.close()
def test_direct_store_waits_for_domain_lock_released_before_deadline(tmp_path: Path) -> None:
config = HostWeightRuntimeConfig(mode=RuntimeMode.PREFERRED, domain=_domain(tmp_path))
assert config.domain is not None
initial = HostWeightRuntime.from_config(config)
assert initial.store is not None
lock_path = tmp_path / "locks" / "domain-init.lock"
with FileLock(lock_path, exclusive=True, deadline=None) as owner:
release = threading.Timer(0.1, owner.close)
release.start()
try:
store = FilesystemHostWeightStore(config.domain, config.capacity, config.integrity, wait=WaitPolicy(5))
finally:
release.join(5)
assert store.root == tmp_path
def test_direct_store_domain_timeout_is_typed(tmp_path: Path) -> None:
config = HostWeightRuntimeConfig(mode=RuntimeMode.PREFERRED, domain=_domain(tmp_path))
assert config.domain is not None
HostWeightRuntime.from_config(config)
lock_path = tmp_path / "locks" / "domain-init.lock"
with FileLock(lock_path, exclusive=True, deadline=None):
with pytest.raises(HostWeightError) as caught:
FilesystemHostWeightStore(config.domain, config.capacity, config.integrity, wait=WaitPolicy(0.01))
assert caught.value.failure.code is FailureCode.DOMAIN_UNAVAILABLE
assert caught.value.failure.retryable
assert lock_is_active(lock_path)
assert not lock_is_active(lock_path)
@pytest.mark.parametrize(
("stage", "code"),
[
@@ -23,7 +23,7 @@ import regex as re
import torch
from safetensors import SafetensorError, safe_open
from ..config import CapacityPolicy, IntegrityPolicy, StorageClass, StorageDomainPolicy, ValidationLevel
from ..config import CapacityPolicy, IntegrityPolicy, StorageClass, StorageDomainPolicy, ValidationLevel, WaitPolicy
from ..errors import FailureCode, HostWeightError, HostWeightFailure, ResolutionStage
from ..identity import CanonicalJson, WeightArtifactIdentity, canonical_json
from ..inspection import (
@@ -429,6 +429,8 @@ class FilesystemHostWeightStore:
domain: StorageDomainPolicy,
capacity: CapacityPolicy,
integrity: IntegrityPolicy,
*,
wait: WaitPolicy = WaitPolicy(),
) -> None:
self.domain_policy = domain
self.capacity_policy = capacity
@@ -464,7 +466,7 @@ class FilesystemHostWeightStore:
)
)
try:
self._initialize_domain()
self._initialize_domain(deadline=time.monotonic() + wait.coordination_timeout_seconds)
except HostWeightError:
raise
except OSError as exc:
@@ -485,7 +487,7 @@ class FilesystemHostWeightStore:
)
) from exc
def _initialize_domain(self) -> None:
def _initialize_domain(self, *, deadline: float) -> None:
try:
self.filesystem_type = detect_filesystem_type(self.root)
detected_storage_class = _storage_class_for_filesystem_type(self.filesystem_type)
@@ -545,7 +547,7 @@ class FilesystemHostWeightStore:
)
)
with FileLock(self.locks_dir / "domain-init.lock", exclusive=True, deadline=None):
with FileLock(self.locks_dir / "domain-init.lock", exclusive=True, deadline=deadline):
domain_path = self.root / _DOMAIN_FILE
if domain_path.exists():
domain_metadata = _read_json_file(domain_path)
+1 -1
View File
@@ -70,7 +70,7 @@ class HostWeightRuntime:
from .filesystem import FilesystemHostWeightStore # noqa: PLC0415
try:
store = FilesystemHostWeightStore(config.domain, config.capacity, config.integrity)
store = FilesystemHostWeightStore(config.domain, config.capacity, config.integrity, wait=config.wait)
except HostWeightError as exc:
return cls(
config,