async def touch_command(
*,
lifetime: Optional[int],
insights: references.Insights,
identity: Identity,
settings: configuration.OperatorSettings,
) -> None:
await asyncio.wait({
insights.ready_namespaces.wait(),
insights.ready_resources.wait(),
})
selector = guess_selector(settings=settings)
resource = insights.backbone.get(selector) if selector else None
if resource is None:
raise RuntimeError(f"Cannot find the peering resource {selector}.")
await aiotasks.wait({
aiotasks.create_guarded_task(
name="peering command", finishable=True, logger=logger,
coro=touch(
namespace=namespace,
resource=resource,
identity=identity,
settings=settings,
lifetime=lifetime),
)
for namespace in insights.namespaces
})
async def test_guard_escalates_on_failure(finishable):
task = create_guarded_task(coro=fail("boo!"),
name='this task',
finishable=finishable)
await asyncio.wait([task], timeout=0.01) # let it start & react
with pytest.raises(Error):
await task
async def test_guard_logs_on_cancellation(assert_logs, caplog):
caplog.set_level(0)
logger = logging.getLogger()
task = create_guarded_task(coro=delay(1), name='this task', logger=logger)
await asyncio.wait([task], timeout=0.01) # let it start
task.cancel()
await asyncio.wait([task], timeout=0.01) # let it react
assert_logs(["This task is cancelled"])
async def test_guard_logs_on_failure(assert_logs, caplog):
caplog.set_level(0)
logger = logging.getLogger()
task = create_guarded_task(coro=fail("boo!"),
name='this task',
logger=logger)
await asyncio.wait([task], timeout=0.01) # let it start & react
assert_logs(["This task has failed: boo!"])
async def test_guard_escalates_on_cancellation(cancellable):
task = create_guarded_task(coro=delay(1),
name='this task',
cancellable=cancellable)
await asyncio.wait([task], timeout=0.01) # let it start
task.cancel()
await asyncio.wait([task], timeout=0.01) # let it react
with pytest.raises(asyncio.CancelledError):
await task
async def test_guard_waits_for_the_flag():
flag = asyncio.Event()
task = create_guarded_task(coro=sample(), name='this task', flag=flag)
await asyncio.wait([task], timeout=0.01) # let it start
assert not task.done()
flag.set()
await asyncio.wait([task], timeout=0.01) # let it react
assert task.done()
async def spawn_missing_peerings(
*,
settings: configuration.OperatorSettings,
identity: peering.Identity,
resources: Collection[references.Resource],
namespaces: Collection[references.Namespace],
ensemble: Ensemble,
) -> None:
for resource, namespace in itertools.product(resources, namespaces):
dkey = EnsembleKey(resource=resource, namespace=namespace)
if dkey not in ensemble.peering_tasks:
what = f"{settings.peering.name}@{namespace}"
is_preactivated = settings.peering.mandatory
conflicts_found = await ensemble.operator_paused.make_toggle(
is_preactivated, name=what)
ensemble.conflicts_found[dkey] = conflicts_found
ensemble.pinging_tasks[dkey] = aiotasks.create_guarded_task(
name=f"peering keep-alive for {what}",
logger=logger,
cancellable=True,
coro=peering.keepalive(namespace=namespace,
resource=resource,
settings=settings,
identity=identity))
ensemble.peering_tasks[dkey] = aiotasks.create_guarded_task(
name=f"peering observer for {what}",
logger=logger,
cancellable=True,
coro=queueing.watcher(settings=settings,
resource=resource,
namespace=namespace,
processor=functools.partial(
peering.process_peering_event,
conflicts_found=conflicts_found,
namespace=namespace,
resource=resource,
settings=settings,
identity=identity)))
# Ensure that all guarded tasks got control for a moment to enter the guard.
await asyncio.sleep(0)
async def test_guard_is_silent_when_cancellable(assert_logs, caplog):
caplog.set_level(0)
logger = logging.getLogger()
task = create_guarded_task(coro=delay(1),
name='this task',
logger=logger,
cancellable=True)
await asyncio.wait([task], timeout=0.01) # let it start
task.cancel()
await asyncio.wait([task], timeout=0.01) # let it react
assert_logs([], prohibited=["This task is cancelled"])
assert not caplog.messages
async def spawn_missing_watchers(
*,
processor: queueing.WatchStreamProcessor,
settings: configuration.OperatorSettings,
resources: Collection[references.Resource],
indexable: Collection[references.Resource],
namespaces: Collection[references.Namespace],
ensemble: Ensemble,
) -> None:
# Block the operator globally until specialised per-resource-kind blockers are created.
# NB: Must be created before the point of parallelisation!
operator_blocked = await ensemble.operator_indexed.make_toggle(
name="orchestration blocker")
# Spawn watchers and create the specialised per-resource-kind blockers.
for resource, namespace in itertools.product(resources, namespaces):
namespace = namespace if resource.namespaced else None
dkey = EnsembleKey(resource=resource, namespace=namespace)
if dkey not in ensemble.watcher_tasks:
what = f"{resource}@{namespace}"
resource_indexed: Optional[primitives.Toggle] = None
if resource in indexable:
resource_indexed = await ensemble.operator_indexed.make_toggle(
name=what)
ensemble.watcher_tasks[dkey] = aiotasks.create_guarded_task(
name=f"watcher for {what}",
logger=logger,
cancellable=True,
coro=queueing.watcher(
operator_paused=ensemble.operator_paused,
operator_indexed=ensemble.operator_indexed,
resource_indexed=resource_indexed,
settings=settings,
resource=resource,
namespace=namespace,
processor=functools.partial(processor, resource=resource)))
# Unblock globally, let the specialised per-resource-kind blockers hold the readiness.
await ensemble.operator_indexed.drop_toggle(operator_blocked)
# Ensure that all guarded tasks got control for a moment to enter the guard.
await asyncio.sleep(0)
async def spawn_tasks(
*,
lifecycle: Optional[lifecycles.LifeCycleFn] = None,
indexers: Optional[indexing.OperatorIndexers] = None,
registry: Optional[registries.OperatorRegistry] = None,
settings: Optional[configuration.OperatorSettings] = None,
memories: Optional[containers.ResourceMemories] = None,
insights: Optional[references.Insights] = None,
identity: Optional[peering.Identity] = None,
standalone: Optional[bool] = None,
priority: Optional[int] = None,
peering_name: Optional[str] = None,
liveness_endpoint: Optional[str] = None,
clusterwide: bool = False,
namespaces: Collection[references.NamespacePattern] = (),
namespace: Optional[references.NamespacePattern] = None, # deprecated
stop_flag: Optional[primitives.Flag] = None,
ready_flag: Optional[primitives.Flag] = None,
vault: Optional[credentials.Vault] = None,
memo: Optional[ephemera.AnyMemo] = None,
_command: Optional[Coroutine[None, None, None]] = None,
) -> Collection[aiotasks.Task]:
"""
Spawn all the tasks needed to run the operator.
The tasks are properly inter-connected with the synchronisation primitives.
"""
loop = asyncio.get_running_loop()
if namespaces and namespace:
raise TypeError("Either namespaces= or namespace= can be passed. Got both.")
elif namespace:
warnings.warn("namespace= is deprecated; use namespaces=[...]", DeprecationWarning)
namespaces = [namespace]
if clusterwide and namespaces:
raise TypeError("The operator can be either cluster-wide or namespaced, not both.")
if not clusterwide and not namespaces:
warnings.warn("Absence of either namespaces or cluster-wide flag will become an error soon."
" For now, switching to the cluster-wide mode for backward compatibility.",
FutureWarning)
clusterwide = True
# All tasks of the operator are synced via these primitives and structures:
lifecycle = lifecycle if lifecycle is not None else lifecycles.get_default_lifecycle()
registry = registry if registry is not None else registries.get_default_registry()
settings = settings if settings is not None else configuration.OperatorSettings()
memories = memories if memories is not None else containers.ResourceMemories()
indexers = indexers if indexers is not None else indexing.OperatorIndexers()
insights = insights if insights is not None else references.Insights()
identity = identity if identity is not None else peering.detect_own_id(manual=False)
vault = vault if vault is not None else credentials.Vault()
memo = memo if memo is not None else ephemera.Memo()
event_queue: posting.K8sEventQueue = asyncio.Queue()
signal_flag: aiotasks.Future = asyncio.Future()
started_flag: asyncio.Event = asyncio.Event()
operator_paused = primitives.ToggleSet(any)
tasks: MutableSequence[aiotasks.Task] = []
# Map kwargs into the settings object.
settings.peering.clusterwide = clusterwide
if peering_name is not None:
settings.peering.mandatory = True
settings.peering.name = peering_name
if standalone is not None:
settings.peering.standalone = standalone
if priority is not None:
settings.peering.priority = priority
# Prepopulate indexers with empty indices -- to be available startup handlers.
indexers.ensure(registry._resource_indexing.get_all_handlers())
# Global credentials store for this operator, also for CRD-reading & peering mode detection.
auth.vault_var.set(vault)
# Special case: pass the settings container through the user-side handlers (no explicit args).
# Toolkits have to keep the original operator context somehow, and the only way is contextvars.
posting.settings_var.set(settings)
# Few common background forever-running infrastructural tasks (irregular root tasks).
tasks.append(aiotasks.create_task(
name="stop-flag checker",
coro=_stop_flag_checker(
signal_flag=signal_flag,
stop_flag=stop_flag)))
tasks.append(aiotasks.create_task(
name="ultimate termination",
coro=_ultimate_termination(
settings=settings,
stop_flag=stop_flag)))
tasks.append(aiotasks.create_task(
name="startup/cleanup activities",
coro=_startup_cleanup_activities(
root_tasks=tasks, # used as a "live" view, populated later.
ready_flag=ready_flag,
started_flag=started_flag,
registry=registry,
settings=settings,
indices=indexers.indices,
vault=vault,
memo=memo))) # to purge & finalize the caches in the end.
# Kill all the daemons gracefully when the operator exits (so that they are not "hung").
tasks.append(aiotasks.create_guarded_task(
name="daemon killer", flag=started_flag, logger=logger,
coro=daemons.daemon_killer(
settings=settings,
memories=memories,
operator_paused=operator_paused)))
# Keeping the credentials fresh and valid via the authentication handlers on demand.
tasks.append(aiotasks.create_guarded_task(
name="credentials retriever", flag=started_flag, logger=logger,
coro=activities.authenticator(
registry=registry,
settings=settings,
indices=indexers.indices,
vault=vault,
memo=memo)))
# K8s-event posting. Events are queued in-memory and posted in the background.
# NB: currently, it is a global task, but can be made per-resource or per-object.
tasks.append(aiotasks.create_guarded_task(
name="poster of events", flag=started_flag, logger=logger,
coro=posting.poster(
backbone=insights.backbone,
event_queue=event_queue)))
# Liveness probing -- so that Kubernetes would know that the operator is alive.
if liveness_endpoint:
tasks.append(aiotasks.create_guarded_task(
name="health reporter", flag=started_flag, logger=logger,
coro=probing.health_reporter(
registry=registry,
settings=settings,
endpoint=liveness_endpoint,
indices=indexers.indices,
memo=memo)))
# Permanent observation of what resource kinds and namespaces are available in the cluster.
# Spawn and cancel dimensional tasks as they come and go; dimensions = resources x namespaces.
tasks.append(aiotasks.create_guarded_task(
name="resource observer", flag=started_flag, logger=logger,
coro=observation.resource_observer(
insights=insights,
registry=registry,
settings=settings)))
tasks.append(aiotasks.create_guarded_task(
name="namespace observer", flag=started_flag, logger=logger,
coro=observation.namespace_observer(
clusterwide=clusterwide,
namespaces=namespaces,
insights=insights,
settings=settings)))
# Explicit command is a hack for the CLI to run coroutines in an operator-like environment.
# If not specified, then use the normal resource processing. It is not exposed publicly (yet).
if _command is not None:
tasks.append(aiotasks.create_guarded_task(
name="the command", flag=started_flag, logger=logger, finishable=True,
coro=_command))
else:
tasks.append(aiotasks.create_guarded_task(
name="multidimensional multitasker", flag=started_flag, logger=logger,
coro=orchestration.ochestrator(
settings=settings,
insights=insights,
identity=identity,
operator_paused=operator_paused,
processor=functools.partial(processing.process_resource_event,
lifecycle=lifecycle,
registry=registry,
settings=settings,
indexers=indexers,
memories=memories,
memobase=memo,
event_queue=event_queue))))
# Ensure that all guarded tasks got control for a moment to enter the guard.
await asyncio.sleep(0)
# On Ctrl+C or pod termination, cancel all tasks gracefully.
if threading.current_thread() is threading.main_thread():
# Handle NotImplementedError when ran on Windows since asyncio only supports Unix signals
try:
loop.add_signal_handler(signal.SIGINT, signal_flag.set_result, signal.SIGINT)
loop.add_signal_handler(signal.SIGTERM, signal_flag.set_result, signal.SIGTERM)
except NotImplementedError:
logger.warning("OS signals are ignored: can't add signal handler in Windows.")
else:
logger.warning("OS signals are ignored: running not in the main thread.")
return tasks
async def spawn_tasks(
*,
lifecycle: Optional[lifecycles.LifeCycleFn] = None,
registry: Optional[registries.OperatorRegistry] = None,
settings: Optional[configuration.OperatorSettings] = None,
memories: Optional[containers.ResourceMemories] = None,
standalone: Optional[bool] = None,
priority: Optional[int] = None,
peering_name: Optional[str] = None,
liveness_endpoint: Optional[str] = None,
namespace: Optional[str] = None,
stop_flag: Optional[primitives.Flag] = None,
ready_flag: Optional[primitives.Flag] = None,
vault: Optional[credentials.Vault] = None,
) -> Collection[aiotasks.Task]:
"""
Spawn all the tasks needed to run the operator.
The tasks are properly inter-connected with the synchronisation primitives.
"""
loop = asyncio.get_running_loop()
# The freezer and the registry are scoped to this whole task-set, to sync them all.
lifecycle = lifecycle if lifecycle is not None else lifecycles.get_default_lifecycle(
)
registry = registry if registry is not None else registries.get_default_registry(
)
settings = settings if settings is not None else configuration.OperatorSettings(
)
memories = memories if memories is not None else containers.ResourceMemories(
)
vault = vault if vault is not None else global_vault
vault = vault if vault is not None else credentials.Vault()
event_queue: posting.K8sEventQueue = asyncio.Queue()
freeze_name = f"{peering_name!r}@{namespace}" if namespace else f"cluster-wide {peering_name!r}"
freeze_checker = primitives.ToggleSet()
freeze_toggle = await freeze_checker.make_toggle(name=freeze_name)
signal_flag: aiotasks.Future = asyncio.Future()
started_flag: asyncio.Event = asyncio.Event()
tasks: MutableSequence[aiotasks.Task] = []
# Map kwargs into the settings object.
if peering_name is not None:
settings.peering.mandatory = True
settings.peering.name = peering_name
if standalone is not None:
settings.peering.standalone = standalone
if priority is not None:
settings.peering.priority = priority
# Global credentials store for this operator, also for CRD-reading & peering mode detection.
auth.vault_var.set(vault)
# Special case: pass the settings container through the user-side handlers (no explicit args).
# Toolkits have to keep the original operator context somehow, and the only way is contextvars.
posting.settings_var.set(settings)
# Few common background forever-running infrastructural tasks (irregular root tasks).
tasks.append(
aiotasks.create_task(name="stop-flag checker",
coro=_stop_flag_checker(signal_flag=signal_flag,
stop_flag=stop_flag)))
tasks.append(
aiotasks.create_task(name="ultimate termination",
coro=_ultimate_termination(settings=settings,
stop_flag=stop_flag)))
tasks.append(
aiotasks.create_task(
name="startup/cleanup activities",
coro=_startup_cleanup_activities(
root_tasks=tasks, # used as a "live" view, populated later.
ready_flag=ready_flag,
started_flag=started_flag,
registry=registry,
settings=settings,
vault=vault))) # to purge & finalize the caches in the end.
# Kill all the daemons gracefully when the operator exits (so that they are not "hung").
tasks.append(
aiotasks.create_guarded_task(
name="daemon killer",
flag=started_flag,
logger=logger,
coro=daemons.daemon_killer(settings=settings, memories=memories)))
# Keeping the credentials fresh and valid via the authentication handlers on demand.
tasks.append(
aiotasks.create_guarded_task(name="credentials retriever",
flag=started_flag,
logger=logger,
coro=activities.authenticator(
registry=registry,
settings=settings,
vault=vault)))
# K8s-event posting. Events are queued in-memory and posted in the background.
# NB: currently, it is a global task, but can be made per-resource or per-object.
tasks.append(
aiotasks.create_guarded_task(
name="poster of events",
flag=started_flag,
logger=logger,
coro=posting.poster(event_queue=event_queue)))
# Liveness probing -- so that Kubernetes would know that the operator is alive.
if liveness_endpoint:
tasks.append(
aiotasks.create_guarded_task(name="health reporter",
flag=started_flag,
logger=logger,
coro=probing.health_reporter(
registry=registry,
settings=settings,
endpoint=liveness_endpoint)))
# Monitor the peers, unless explicitly disabled.
if await peering.detect_presence(namespace=namespace, settings=settings):
identity = peering.detect_own_id(manual=False)
tasks.append(
aiotasks.create_guarded_task(name="peering keepalive",
flag=started_flag,
logger=logger,
coro=peering.keepalive(
namespace=namespace,
settings=settings,
identity=identity)))
tasks.append(
aiotasks.create_guarded_task(
name="watcher of peering",
flag=started_flag,
logger=logger,
coro=queueing.watcher(
namespace=namespace,
settings=settings,
resource=peering.guess_resource(namespace=namespace),
processor=functools.partial(peering.process_peering_event,
namespace=namespace,
settings=settings,
identity=identity,
freeze_toggle=freeze_toggle))))
# Resource event handling, only once for every known resource (de-duplicated).
for resource in registry.resources:
tasks.append(
aiotasks.create_guarded_task(
name=f"watcher of {resource.name}",
flag=started_flag,
logger=logger,
coro=queueing.watcher(namespace=namespace,
settings=settings,
resource=resource,
freeze_checker=freeze_checker,
processor=functools.partial(
processing.process_resource_event,
lifecycle=lifecycle,
registry=registry,
settings=settings,
memories=memories,
resource=resource,
event_queue=event_queue))))
# On Ctrl+C or pod termination, cancel all tasks gracefully.
if threading.current_thread() is threading.main_thread():
# Handle NotImplementedError when ran on Windows since asyncio only supports Unix signals
try:
loop.add_signal_handler(signal.SIGINT, signal_flag.set_result,
signal.SIGINT)
loop.add_signal_handler(signal.SIGTERM, signal_flag.set_result,
signal.SIGTERM)
except NotImplementedError:
logger.warning(
"OS signals are ignored: can't add signal handler in Windows.")
else:
logger.warning(
"OS signals are ignored: running not in the main thread.")
return tasks
