def list_metrics(experiment_id: Optional[str] = None,
operator_id: Optional[str] = None,
run_id: Optional[str] = None) -> List[Dict[str, object]]:
"""Lists metrics from object storage.
Args:
experiment_id (str, optional): the experiment uuid. Defaults to None.
operator_id (str, optional): the operator uuid. Defaults to None.
run_id (str, optional): the run id. Defaults to None.
Returns:
list: A list of metrics.
Raises:
TypeError: when experiment_id is undefined in args and env.
TypeError: when operator_id is undefined in args and env.
"""
if experiment_id is None:
experiment_id = get_experiment_id()
if operator_id is None:
operator_id = get_operator_id()
# ensures MinIO bucket exists
make_bucket(BUCKET_NAME)
if run_id is None:
# gets run_id from env variable
# Attention: returns None if env is unset
run_id = get_run_id()
elif run_id == "latest":
try:
metadata = stat_metadata(experiment_id, operator_id)
run_id = metadata.get("run_id")
except FileNotFoundError:
return []
try:
object_name = operator_filepath(METRICS_FILE, experiment_id, operator_id, run_id)
data = MINIO_CLIENT.get_object(
bucket_name=BUCKET_NAME,
object_name=object_name,
)
except (NoSuchBucket, NoSuchKey):
raise FileNotFoundError(f"No such file or directory: '{experiment_id}'")
return load(data)
def save_metrics(experiment_id: Optional[str] = None,
operator_id: Optional[str] = None,
run_id: Optional[str] = None,
**kwargs):
"""Saves metrics of an experiment to the object storage.
Args:
experiment_id (str, optional): the experiment uuid. Defaults to None
operator_id (str, optional): the operator uuid. Defaults to None
run_id (str, optional): the run id. Defaults to None.
**kwargs: the metrics dict.
Raises:
TypeError: when experiment_id is undefined in args and env.
TypeError: when operator_id is undefined in args and env.
"""
if experiment_id is None:
experiment_id = get_experiment_id()
if operator_id is None:
operator_id = get_operator_id()
if run_id is None:
# gets run_id from env variables
# Attention: returns None if env is unset
run_id = get_run_id()
# ensures MinIO bucket exists
make_bucket(BUCKET_NAME)
if run_id:
metadata = {}
try:
metadata = stat_metadata(experiment_id, operator_id)
if run_id == "latest":
run_id = metadata.get("run_id")
except FileNotFoundError:
pass
metadata["run_id"] = run_id
# encodes metadata to JSON format and uploads to MinIO
buffer = BytesIO(dumps(metadata).encode())
MINIO_CLIENT.put_object(
bucket_name=BUCKET_NAME,
object_name=f'experiments/{experiment_id}/operators/{operator_id}/.metadata',
data=buffer,
length=buffer.getbuffer().nbytes,
)
object_name = operator_filepath(METRICS_FILE, experiment_id, operator_id, run_id)
encoded_metrics = []
# retrieves the metrics saved previosuly
try:
data = MINIO_CLIENT.get_object(
bucket_name=BUCKET_NAME,
object_name=object_name,
)
encoded_metrics = loads(data.read())
except NoSuchKey:
pass
# appends new metrics
encoded_metrics.extend(_encode_metrics(kwargs))
# puts metrics into buffer
buffer = BytesIO(dumps(encoded_metrics).encode())
# uploads metrics to MinIO
MINIO_CLIENT.put_object(
bucket_name=BUCKET_NAME,
object_name=object_name,
data=buffer,
length=buffer.getbuffer().nbytes,
)
def stat_dataset(name: str,
run_id: Optional[str] = None,
operator_id: Optional[str] = None) -> Dict[str, str]:
"""Retrieves the metadata of a dataset.
Args:
name (str): the dataset name.
run_id (str, optional): the run id of trainning pipeline. Defaults to None.
operator_id (str, optional): the operator uuid. Defaults to None.
Returns:
dict: The metadata.
Raises:
FileNotFoundError: If dataset does not exist in the object storage.
"""
metadata = {}
# remove /tmp/data/ from dataset name
# because in jupyter we use dataset with full path
name = name.replace("/tmp/data/", "")
if run_id == "latest":
metadata = stat_dataset(name)
run_id = metadata.get("run_id")
if run_id is None:
# gets run_id from env variables
# Attention: returns None if env is unset
run_id = get_run_id()
if run_id and operator_id:
# get metadata for a specific operator of a run, if exists
object_name = _metadata_filepath(name, run_id, operator_id)
elif run_id:
# if no metadata was generated by the operator,
# get the last one generated by the pipeline flow
object_name = _metadata_filepath(name, run_id)
if not metadata_exists(name, run_id):
# if it is at the beginning of a run,
# there will be no metadata generated by run_id
object_name = _metadata_filepath(name)
else:
# unable to get run_id automatically,
# this function is probably being called out of a run
object_name = _metadata_filepath(name)
else:
# get path according to received parameters
run_id = None if run_id == "root" else run_id
object_name = _metadata_filepath(name, run_id, operator_id)
try:
# reads the .metadata file
data = MINIO_CLIENT.get_object(
bucket_name=BUCKET_NAME,
object_name=object_name,
)
# decodes the metadata (which is in JSON format)
metadata = loads(data.read())
except (NoSuchBucket, NoSuchKey):
raise FileNotFoundError("The specified dataset does not exist")
return metadata
def load_dataset(
name: str,
run_id: Optional[str] = None,
operator_id: Optional[str] = None,
page: Optional[int] = None,
page_size: Optional[int] = None) -> Union[pd.DataFrame, BinaryIO]:
"""Retrieves the contents of a dataset.
If run_id exists, then loads the dataset from the specified run.
If the dataset does not exist for given run_id/operator_id return the
'original' dataset
Args:
name (str): the dataset name.
run_id (str, optional): the run id of training pipeline. Defaults to None.
operator_id (str, optional): the operator uuid. Defaults to None.
Returns:
The contents of a dataset. Either a `pandas.DataFrame` or an `BinaryIO` buffer.
Raises:
FileNotFoundError: If dataset does not exist in the object storage.
"""
if run_id is None:
# gets run_id from env variable
# Attention: returns None if env is unset
run_id = get_run_id()
elif run_id == "latest":
metadata = stat_dataset(name)
run_id = metadata.get("run_id")
# when the dataset does not exist for given run_id/operator_id
# must return the 'original' dataset
# unset run_id so data_filepath points to the 'original' dataset
if run_id and operator_id:
try:
metadata = stat_dataset(name, run_id, operator_id)
except FileNotFoundError:
run_id = None
elif run_id:
try:
run_metadata = stat_dataset(name, run_id)
operator_id = run_metadata.get("operator_id")
except FileNotFoundError:
run_id = None
# builds the path to the dataset file
path = _data_filepath(name, run_id, operator_id)
if page_size and page_size > 0:
nrows = page_size
else:
nrows = None
if page and page > 0:
skiprows = (page - 1) * page_size
else:
skiprows = None
try:
metadata = stat_dataset(name, run_id, operator_id)
dataset = pd.read_csv(S3FS.open(path),
header=0,
index_col=False,
nrows=nrows,
skiprows=skiprows)
dtypes = dict((column, "object") for column, ftype in zip(
metadata["columns"], metadata["featuretypes"])
if ftype in [CATEGORICAL, DATETIME])
dataset = dataset.astype(dtypes)
except (UnicodeDecodeError, pd.errors.EmptyDataError,
pd.errors.ParserError):
# reads the raw file
data = MINIO_CLIENT.get_object(
bucket_name=BUCKET_NAME,
object_name=path.lstrip(f"{BUCKET_NAME}/"),
)
return BytesIO(data.read())
except KeyError:
# metadata file does not contains "columns" or "featuretypes"
# ignore this error and return dataset without cast its type
pass
except FileNotFoundError:
raise FileNotFoundError("The specified dataset does not exist")
return dataset
def save_dataset(name: str,
data: Union[pd.DataFrame, BinaryIO] = None,
df: pd.DataFrame = None,
metadata: Optional[Dict[str, str]] = None,
run_id: Optional[str] = None,
operator_id: Optional[str] = None):
"""Saves a dataset and its metadata to the object storage.
Args:
name (str): the dataset name.
data (pandas.DataFrame, BinaryIO, optional): the dataset contents as a
pandas.DataFrame or an `BinaryIO` buffer. Defaults to None.
df (pandas.DataFrame, optional): the dataset contents as an `pandas.DataFrame`.
df exists only for compatibility with existing components.
Use "data" for all types of datasets. Defaults to None.
metadata (dict, optional): metadata about the dataset. Defaults to None..
run_id (str, optional): the run id. Defaults to None.
operator_id (str, optional): the operator uuid. Defaults to None.
Raises:
PermissionError: If dataset was read only.
"""
# ensures MinIO bucket exists
make_bucket(BUCKET_NAME)
if run_id is None:
# gets run_id from env variables
# Attention: returns None if env is unset
run_id = get_run_id()
if operator_id is None:
# gets operator_id from env variables
# Attention: returns None if env is unset
operator_id = get_operator_id(raise_for_none=False)
# df exists only for compatibility with existing components
# from now on one must use "data" for all types of datasets
if df is not None:
data = df
try:
# gets metadata (if dataset exists)
stored_metadata = stat_dataset(name, run_id)
metadata_should_be_updated = False
# update stored metadata values
if metadata:
stored_metadata.update(metadata)
elif isinstance(data, pd.DataFrame):
metadata_should_be_updated = True
metadata = stored_metadata
except FileNotFoundError:
metadata_should_be_updated = False
# builds metadata dict:
# sets filename and run_id
if metadata is None:
metadata = {}
metadata["filename"] = name
if isinstance(data, pd.DataFrame):
# sets metadata specific for pandas.DataFrame:
# columns, featuretypes
metadata["columns"] = data.columns.tolist()
metadata["total"] = len(data.index)
if "featuretypes" not in metadata:
metadata["featuretypes"] = infer_featuretypes(data)
# if the metadata was given (set manually), ignore updates, otherwise
# search for changes and then update current featuretypes to be even with columns
if metadata_should_be_updated:
previous_metadata = stat_dataset(name, run_id)
previous_columns = previous_metadata["columns"]
previous_featuretypes = previous_metadata["featuretypes"]
column_to_type = dict(zip(previous_columns, previous_featuretypes))
new_featuretypes = []
for new_column in metadata["columns"]:
if new_column in column_to_type:
new_featuretypes.append(column_to_type[new_column])
else:
new_featuretypes.append(
infer_featuretypes(pd.DataFrame(data[new_column]))[0])
metadata["featuretypes"] = new_featuretypes
if run_id:
metadata["run_id"] = run_id
# When saving a dataset of a run, also
# set the run_id in datasets/<name>.metadata
# This enables load_dataset by run="latest"
try:
root_metadata = stat_dataset(name, "root")
except FileNotFoundError:
root_metadata = {}
root_metadata["run_id"] = run_id
object_name = _metadata_filepath(name)
# encodes metadata to JSON format
buffer = BytesIO(dumps(root_metadata).encode())
MINIO_CLIENT.put_object(
bucket_name=BUCKET_NAME,
object_name=object_name,
data=buffer,
length=buffer.getbuffer().nbytes,
)
# create a run metadata to save the last operator id
# to dataset get loaded on next step of the pipeline flow
metadata["operator_id"] = operator_id
object_name = _metadata_filepath(name, run_id=run_id)
buffer = BytesIO(dumps(metadata).encode())
MINIO_CLIENT.put_object(
bucket_name=BUCKET_NAME,
object_name=object_name,
data=buffer,
length=buffer.getbuffer().nbytes,
)
path = _data_filepath(name, run_id, operator_id)
if isinstance(data, pd.DataFrame):
# uploads dataframe to MinIO as a .csv file
temp_file = tempfile.NamedTemporaryFile(dir='.', delete=False)
data.to_csv(temp_file.name, header=True, index=False)
MINIO_CLIENT.fput_object(bucket_name=BUCKET_NAME,
object_name=path.lstrip(f"{BUCKET_NAME}/"),
file_path=temp_file.name)
temp_file.close()
os.remove(temp_file.name)
else:
# uploads raw data to MinIO
buffer = BytesIO(data.read())
MINIO_CLIENT.put_object(
bucket_name=BUCKET_NAME,
object_name=path.lstrip(f"{BUCKET_NAME}/"),
data=buffer,
length=buffer.getbuffer().nbytes,
)
object_name = _metadata_filepath(name, run_id, operator_id)
# encodes metadata to JSON format
buffer = BytesIO(dumps(metadata).encode())
MINIO_CLIENT.put_object(
bucket_name=BUCKET_NAME,
object_name=object_name,
data=buffer,
length=buffer.getbuffer().nbytes,
)