clearml-server/apiserver/bll/task/task_cleanup.py

325 lines
11 KiB
Python

from datetime import datetime
from itertools import chain
from operator import attrgetter
from typing import Sequence, Set, Tuple
import attr
from boltons.iterutils import partition, bucketize, first
from mongoengine import NotUniqueError
from pymongo.errors import DuplicateKeyError
from apiserver.apierrors import errors
from apiserver.bll.event import EventBLL
from apiserver.bll.event.event_bll import PlotFields
from apiserver.bll.task.utils import deleted_prefix
from apiserver.config_repo import config
from apiserver.database.model.model import Model
from apiserver.database.model.task.task import Task, TaskStatus, ArtifactModes
from apiserver.database.model.url_to_delete import (
StorageType,
UrlToDelete,
FileType,
DeletionStatus,
)
from apiserver.database.utils import id as db_id
log = config.logger(__file__)
event_bll = EventBLL()
async_events_delete = config.get("services.tasks.async_events_delete", False)
@attr.s(auto_attribs=True)
class TaskUrls:
model_urls: Sequence[str]
event_urls: Sequence[str]
artifact_urls: Sequence[str]
def __add__(self, other: "TaskUrls"):
if not other:
return self
return TaskUrls(
model_urls=list(set(self.model_urls) | set(other.model_urls)),
event_urls=list(set(self.event_urls) | set(other.event_urls)),
artifact_urls=list(set(self.artifact_urls) | set(other.artifact_urls)),
)
@attr.s(auto_attribs=True)
class CleanupResult:
"""
Counts of objects modified in task cleanup operation
"""
updated_children: int
updated_models: int
deleted_models: int
urls: TaskUrls = None
def __add__(self, other: "CleanupResult"):
if not other:
return self
return CleanupResult(
updated_children=self.updated_children + other.updated_children,
updated_models=self.updated_models + other.updated_models,
deleted_models=self.deleted_models + other.deleted_models,
urls=self.urls + other.urls if self.urls else other.urls,
)
def collect_plot_image_urls(company: str, task_or_model: str) -> Set[str]:
urls = set()
next_scroll_id = None
while True:
events, next_scroll_id = event_bll.get_plot_image_urls(
company_id=company, task_id=task_or_model, scroll_id=next_scroll_id
)
if not events:
break
for event in events:
event_urls = event.get(PlotFields.source_urls)
if event_urls:
urls.update(set(event_urls))
return urls
def collect_debug_image_urls(company: str, task_or_model: str) -> Set[str]:
"""
Return the set of unique image urls
Uses DebugImagesIterator to make sure that we do not retrieve recycled urls
"""
after_key = None
urls = set()
while True:
res, after_key = event_bll.get_debug_image_urls(
company_id=company, task_id=task_or_model, after_key=after_key,
)
urls.update(res)
if not after_key:
break
return urls
supported_storage_types = {
"https://": StorageType.fileserver,
"http://": StorageType.fileserver,
}
def _schedule_for_delete(
company: str, user: str, task_id: str, urls: Set[str], can_delete_folders: bool,
) -> Set[str]:
urls_per_storage = bucketize(
urls,
key=lambda u: first(
type_
for prefix, type_ in supported_storage_types.items()
if u.startswith(prefix)
),
)
urls_per_storage.pop(None, None)
processed_urls = set()
for storage_type, storage_urls in urls_per_storage.items():
delete_folders = (storage_type == StorageType.fileserver) and can_delete_folders
scheduled_to_delete = set()
for url in storage_urls:
folder = None
if delete_folders:
folder, _, _ = url.rpartition("/")
to_delete = folder or url
if to_delete in scheduled_to_delete:
processed_urls.add(url)
continue
try:
UrlToDelete(
id=db_id(),
company=company,
user=user,
url=to_delete,
task=task_id,
created=datetime.utcnow(),
storage_type=storage_type,
type=FileType.folder if folder else FileType.file,
).save()
except (DuplicateKeyError, NotUniqueError):
existing = UrlToDelete.objects(company=company, url=to_delete).first()
if existing:
existing.update(
user=user,
task=task_id,
created=datetime.utcnow(),
retry_count=0,
unset__last_failure_time=1,
unset__last_failure_reason=1,
status=DeletionStatus.created,
)
processed_urls.add(url)
scheduled_to_delete.add(to_delete)
return processed_urls
def cleanup_task(
company: str,
user: str,
task: Task,
force: bool = False,
update_children=True,
return_file_urls=False,
delete_output_models=True,
delete_external_artifacts=True,
) -> CleanupResult:
"""
Validate task deletion and delete/modify all its output.
:param task: task object
:param force: whether to delete task with published outputs
:return: count of delete and modified items
"""
published_models, draft_models, in_use_model_ids = verify_task_children_and_ouptuts(
task, force
)
delete_external_artifacts = delete_external_artifacts and config.get(
"services.async_urls_delete.enabled", False
)
event_urls, artifact_urls, model_urls = set(), set(), set()
if return_file_urls or delete_external_artifacts:
event_urls = collect_debug_image_urls(task.company, task.id)
event_urls.update(collect_plot_image_urls(task.company, task.id))
if task.execution and task.execution.artifacts:
artifact_urls = {
a.uri
for a in task.execution.artifacts.values()
if a.mode == ArtifactModes.output and a.uri
}
model_urls = {
m.uri for m in draft_models if m.uri and m.id not in in_use_model_ids
}
deleted_task_id = f"{deleted_prefix}{task.id}"
updated_children = 0
if update_children:
updated_children = Task.objects(parent=task.id).update(parent=deleted_task_id)
deleted_models = 0
updated_models = 0
for models, allow_delete in ((draft_models, True), (published_models, False)):
if not models:
continue
if delete_output_models and allow_delete:
model_ids = set(m.id for m in models if m.id not in in_use_model_ids)
for m_id in model_ids:
if return_file_urls or delete_external_artifacts:
event_urls.update(collect_debug_image_urls(task.company, m_id))
event_urls.update(collect_plot_image_urls(task.company, m_id))
try:
event_bll.delete_task_events(
task.company,
m_id,
allow_locked=True,
model=True,
async_delete=async_events_delete,
)
except errors.bad_request.InvalidModelId as ex:
log.info(f"Error deleting events for the model {m_id}: {str(ex)}")
deleted_models += Model.objects(id__in=list(model_ids)).delete()
if in_use_model_ids:
Model.objects(id__in=list(in_use_model_ids)).update(unset__task=1)
continue
if update_children:
updated_models += Model.objects(id__in=[m.id for m in models]).update(
task=deleted_task_id
)
else:
Model.objects(id__in=[m.id for m in models]).update(unset__task=1)
event_bll.delete_task_events(
task.company, task.id, allow_locked=force, async_delete=async_events_delete
)
if delete_external_artifacts:
scheduled = _schedule_for_delete(
task_id=task.id,
company=company,
user=user,
urls=event_urls | model_urls | artifact_urls,
can_delete_folders=not in_use_model_ids and not published_models,
)
for urls in (event_urls, model_urls, artifact_urls):
urls.difference_update(scheduled)
return CleanupResult(
deleted_models=deleted_models,
updated_children=updated_children,
updated_models=updated_models,
urls=TaskUrls(
event_urls=list(event_urls),
artifact_urls=list(artifact_urls),
model_urls=list(model_urls),
)
if return_file_urls
else None,
)
def verify_task_children_and_ouptuts(
task, force: bool
) -> Tuple[Sequence[Model], Sequence[Model], Set[str]]:
if not force:
published_children_count = Task.objects(
parent=task.id, status=TaskStatus.published
).count()
if published_children_count:
raise errors.bad_request.TaskCannotBeDeleted(
"has children, use force=True",
task=task.id,
children=published_children_count,
)
model_fields = ["id", "ready", "uri"]
published_models, draft_models = partition(
Model.objects(task=task.id).only(*model_fields), key=attrgetter("ready"),
)
if not force and published_models:
raise errors.bad_request.TaskCannotBeDeleted(
"has output models, use force=True",
task=task.id,
models=len(published_models),
)
if task.models and task.models.output:
model_ids = [m.model for m in task.models.output]
for output_model in Model.objects(id__in=model_ids).only(*model_fields):
if output_model.ready:
if not force:
raise errors.bad_request.TaskCannotBeDeleted(
"has output model, use force=True",
task=task.id,
model=output_model.id,
)
published_models.append(output_model)
else:
draft_models.append(output_model)
in_use_model_ids = {}
if draft_models:
model_ids = {m.id for m in draft_models}
dependent_tasks = Task.objects(models__input__model__in=list(model_ids)).only(
"id", "models"
)
in_use_model_ids = model_ids & {
m.model
for m in chain.from_iterable(
t.models.input for t in dependent_tasks if t.models
)
}
return published_models, draft_models, in_use_model_ids