pai/dsw.py (185 lines of code) (raw):
import os
import posixpath
from enum import Enum
from typing import Any, Dict, List, Optional, Union
from .common.logging import get_logger
from .common.oss_utils import OssUriObj, is_oss_uri
from .common.utils import is_dataset_id, is_nas_uri
from .libs.alibabacloud_pai_dsw20220101.models import (
GetInstanceRequest,
GetInstanceResponse,
GetInstanceResponseBody,
UpdateInstanceRequest,
UpdateInstanceRequestDatasets,
)
from .session import get_default_session
logger = get_logger()
class OptionType(str, Enum):
"""
The type of options for mounting data sources in DSW (Data Science Workshop).
This enum defines different mounting modes for OSS (Object Storage Service)
and custom datasets in DSW. Each mode has specific characteristics and use cases:
Attributes:
FastReadWrite: Allows fast read and write operations. Suitable for training
data and models, but may have data consistency issues with concurrent
read/write operations. Not recommended for work directories.
IncrementalReadWrite: Ensures data consistency for incremental writes,
but may have consistency issues when overwriting existing data.
Slightly slower read speed. Suitable for saving training model weights.
ConsistentReadWrite: Maintains data consistency during concurrent read/write
operations. Suitable for scenarios requiring high data consistency
but can tolerate slower read speeds. Ideal for saving code projects.
ReadOnly: Allows only read operations, no writing permitted.
Suitable for mounting public datasets.
The choice of option type affects the underlying Jindo configuration for
mounting OSS data in DSW. Users can select the appropriate mode based on
their specific use case and performance requirements.
"""
FastReadWrite = "FastReadWrite"
IncrementalReadWrite = "IncrementalReadWrite"
ConsistentReadWrite = "ConsistentReadWrite"
ReadOnly = "ReadOnly"
def _default_instance() -> "DswInstance":
"""
Get the default DSW Instance.
Returns:
DswInstance: The default DSW Instance.
"""
instance_id = os.getenv("DSW_INSTANCE_ID")
if not instance_id:
raise RuntimeError(
"Environment variable 'DSW_INSTANCE_ID' is not set, please check if you are running in DSW environment"
)
return DswInstance(instance_id)
def mount(
source: str,
mount_point: str = None,
options: Optional[Dict[str, Any]] = None,
option_type: Optional[OptionType] = None,
) -> str:
"""
Dynamic mount a dataset to the DSW Instance.
Args:
source (str): The source to be mounted, can be a dataset id or an OSS uri.
mount_point (str): Target mount point in the instance, if not specified, the
mount point be generate with given source under the default mount point.
options (dict): Options that apply to when mount a data source, can not be
specified with option_type.
option_type(str): Preset data source mount options, can not be specified with
options.
Returns:
str: The mount point of the data source.
Examples:
>>> mount_point = mount("oss://my-bucket/my-object/path/to/dir")
>>> mount_point = mount("oss://my-bucket/my-object/path/to/dir", mount_point="/my-target", option_type=OptionType.FastReadWrite)
>>> mount_point = mount("oss://my-bucket/my-object/path/to/dir", mount_point="/my-target", options={
"fs.oss.download.thread.concurrency": "8", # CPU core count * 2
"fs.oss.upload.thread.concurrency": "8", # CPU core count * 2
# jindo args, refer to https://help.aliyun.com/zh/emr/emr-on-ecs/user-guide/user-guide-of-jindofuse
"fs.jindo.args": "-oattr_timeout=3 -oentry_timeout=0 -onegative_timeout=0 -oauto_cache -ono_symlink" # jindo args
})
"""
instance = _default_instance()
return instance.mount(
source,
mount_point,
options=options,
option_type=option_type,
)
def unmount(mount_point: str) -> None:
"""
Unmount a dynamic mount dataset from the DSW Instance.
Dynamic mount dataset is a special dataset in DSW, which is used to mount data sources
dynamically when the instance is running. When unmount it, the dataset will be removed
from the instance.
Args:
mount_point (str): The mount point to be unmounted.
Returns:
None
"""
instance = _default_instance()
instance.unmount(mount_point)
def list_dataset_configs(dynamic_only: bool = False) -> List[Dict[str, Any]]:
"""
List all the datasets available in the DSW Instance.
Args:
dynamic_only (bool): Whether to list only the dynamic mount datasets.
Returns:
list: A list of dataset details.
"""
instance = _default_instance()
datasets = instance._get_instance_info().datasets
if dynamic_only:
datasets = [ds for ds in datasets if ds.dynamic]
return [ds.to_map() for ds in datasets]
def default_dynamic_mount_path():
"""Get the default dynamic mount path of the DSW Instance.
Returns:
str: The default dynamic mount path of the DSW Instance.
"""
instance = _default_instance()
return instance.default_dynamic_mount_path()
def get_dynamic_mount_config() -> Dict[str, Any]:
"""
Get the dynamic mount config of the DSW Instance.
Returns:
dict: The dynamic mount config of the DSW Instance.
"""
instance = _default_instance()
return instance.get_dynamic_mount_config()
class DswInstance:
"""A object representing a DSW notebook instance"""
def __init__(self, instance_id: str):
self.instance_id = instance_id
self._instance_info: GetInstanceResponseBody = self._get_instance_info()
def _get_instance_info(self):
session = get_default_session()
resp: GetInstanceResponse = session._acs_dsw_client.get_instance(
self.instance_id,
request=GetInstanceRequest(),
)
return resp.body
def get_dynamic_mount_config(self):
"""Get the dynamic mount config of the DSW Instance.
Returns:
dict: The dynamic mount config of the DSW Instance.
"""
return self._instance_info.dynamic_mount.to_map()
def default_dynamic_mount_path(self) -> Optional[str]:
"""Get the default dynamic mount path of the DSW Instance.
Returns:
str: The default dynamic mount path of the DSW Instance.
"""
if not self._instance_info.dynamic_mount.enable:
return None
if not self._instance_info.dynamic_mount.mount_points:
return "/mnt/dynamic"
return self._instance_info.dynamic_mount.mount_points[0].root_path
def mount(
self,
source: str,
mount_point: str = None,
options: Union[str] = None,
option_type: Union[OptionType] = None,
) -> str:
"""
Dynamic mount a data source to the DSW Instance.
Args:
source (str): The source to be mounted, can be a dataset id or OSS/NAS uri.
mount_point (str): Target mount point in the instance, if not specified, the
mount point be generated with given source under the default mount point.
options (str): Options that apply to when mount a data source, can not be
specified with option_type.
option_type(str): Preset data source mount options, can not be specified with
options.
Returns:
str: The mount point of the data source.
"""
if options and option_type:
raise ValueError(
"options and option_type cannot be specified at the same time"
)
if not self._instance_info.dynamic_mount.enable:
raise RuntimeError(
"Dynamic mount is not enabled for the DSW instance: {}".format(
self.instance_id
)
)
sess = get_default_session()
default_root_path = self.default_dynamic_mount_path()
if is_oss_uri(source):
obj = OssUriObj(source)
if not obj.endpoint:
obj.endpoint = sess.oss_endpoint or sess._get_default_oss_endpoint()
# ensure mount source OSS uri is a directory
_, dir_path, _ = obj.parse_object_key()
uri = f"oss://{obj.bucket_name}.{obj.endpoint}{dir_path}"
dataset_id = None
dataset_version = None
elif is_nas_uri(source):
uri = source
dataset_id = None
dataset_version = None
else:
dataset_id = source
uri = None
if "/" in dataset_id:
dataset_id, dataset_version = dataset_id.split("/", 1)
else:
dataset_version = "v1"
if (
not is_oss_uri(source)
and not is_nas_uri(source)
and not is_dataset_id(source)
):
raise ValueError("Source must be oss uri or nas uri or dataset id")
if not mount_point:
if is_oss_uri(source):
obj = OssUriObj(source)
mount_point = f"{obj.bucket_name}/{obj.object_key}"
elif is_nas_uri(source):
raise ValueError("Mount point is required for nas url.")
else:
mount_point = source
if not posixpath.isabs(mount_point):
mount_point = posixpath.join(default_root_path, mount_point)
resp: GetInstanceResponse = sess._acs_dsw_client.get_instance(
self.instance_id, request=GetInstanceRequest()
)
datasets = [
UpdateInstanceRequestDatasets().from_map(ds.to_map())
for ds in resp.body.datasets
]
datasets.append(
UpdateInstanceRequestDatasets(
dataset_id=dataset_id,
dataset_version=dataset_version,
dynamic=True,
mount_path=mount_point,
option_type=option_type,
options=options,
uri=uri,
)
)
request = UpdateInstanceRequest(
datasets=datasets,
)
update_resp = sess._acs_dsw_client.update_instance(
instance_id=self.instance_id, request=request
)
print("Mount succeed, request id: {}".format(update_resp.body.request_id))
return mount_point
def unmount(self, mount_point: str) -> None:
"""
Unmount a data source from the DSW Instance.
Args:
mount_point (str): The mount point to be unmounted.
Returns:
None
"""
sess = get_default_session()
resp: GetInstanceResponse = sess._acs_dsw_client.get_instance(
self.instance_id, request=GetInstanceRequest()
)
datasets = [
UpdateInstanceRequestDatasets().from_map(ds.to_map())
for ds in resp.body.datasets
]
unmount_ds = [ds for ds in datasets if ds.mount_path == mount_point]
if not unmount_ds:
raise ValueError(f"Not found dataset to unmount: {mount_point}")
if len(unmount_ds) > 1:
raise RuntimeError(f"Found multiple datasets to unmount: {mount_point}")
dataset = unmount_ds[0]
if not dataset.dynamic:
raise ValueError(f"Dataset is not a dynamic mount point: {mount_point}")
request = UpdateInstanceRequest(
datasets=[
UpdateInstanceRequestDatasets().from_map(ds.to_map())
for ds in resp.body.datasets
if ds.mount_path != mount_point
]
)
if not request.datasets:
request.disassociate_datasets = True
update_resp = sess._acs_dsw_client.update_instance(
instance_id=self.instance_id, request=request
)
print("Unmount succeed, request id: {}".format(update_resp.body.request_id))