# Copyright 2023 Iguazio
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import warnings
from datetime import datetime
from typing import Optional, Union
import pandas as pd
from storey import EmitEveryEvent, EmitPolicy
import mlrun
import mlrun.common.schemas
from ..config import config as mlconf
from ..data_types import InferOptions
from ..datastore import get_store_uri
from ..datastore.sources import BaseSourceDriver, source_kind_to_driver
from ..datastore.targets import (
TargetTypes,
get_default_targets,
get_offline_target,
get_online_target,
get_target_driver,
update_targets_run_id_for_ingest,
validate_target_list,
validate_target_placement,
)
from ..features import Entity, Feature
from ..model import (
DataSource,
DataTarget,
DataTargetBase,
ModelObj,
ObjectDict,
ObjectList,
VersionedObjMetadata,
)
from ..runtimes import BaseRuntime
from ..runtimes.function_reference import FunctionReference
from ..serving.states import BaseStep, RootFlowStep, previous_step, queue_class_names
from ..serving.utils import StepToDict
from ..utils import StorePrefix, logger
from .common import RunConfig, verify_feature_set_permissions
aggregates_step = "Aggregates"
class FeatureAggregation(ModelObj):
"""feature aggregation requirements"""
def __init__(
self, name=None, column=None, operations=None, windows=None, period=None
):
self.name = name
self.column = column
self.operations = operations or []
self.windows = windows or []
self.period = period
[docs]class FeatureSetSpec(ModelObj):
def __init__(
self,
owner=None,
description=None,
entities=None,
features=None,
partition_keys=None,
timestamp_key=None,
label_column=None,
relations=None,
source=None,
targets=None,
graph=None,
function=None,
analysis=None,
engine=None,
output_path=None,
passthrough=None,
):
"""Feature set spec object, defines the feature-set's configuration.
.. warning::
This class should not be modified directly. It is managed by the parent feature-set object or using
feature-store APIs. Modifying the spec manually may result in unpredictable behaviour.
:param description: text description (copied from parent feature-set)
:param entities: list of entity (index key) names or :py:class:`~mlrun.features.FeatureSet.Entity`
:param features: list of features - :py:class:`~mlrun.features.FeatureSet.Feature`
:param partition_keys: list of fields to partition results by (other than the default timestamp key)
:param timestamp_key: timestamp column name
:param label_column: name of the label column (the one holding the target (y) values)
:param targets: list of data targets
:param graph: the processing graph
:param function: MLRun runtime to execute the feature-set in
:param engine: name of the processing engine (storey, pandas, or spark), defaults to storey
:param output_path: default location where to store results (defaults to MLRun's artifact path)
:param passthrough: if true, ingest will skip offline targets, and get_offline_features will
read directly from source
"""
self._features: ObjectList = None
self._entities: ObjectList = None
self._targets: ObjectList = None
self._graph: RootFlowStep = None
self._source = None
self._engine = None
self._function: FunctionReference = None
self._relations: ObjectDict = None
self.owner = owner
self.description = description
self.entities: list[Union[Entity, str]] = entities or []
self.relations: dict[str, Union[Entity, str]] = relations or {}
self.features: list[Feature] = features or []
self.partition_keys = partition_keys or []
self.timestamp_key = timestamp_key
self.source = source
self.targets = targets or []
self.graph = graph
self.label_column = label_column
self.function = function
self.analysis = analysis or {}
self.engine = engine
self.output_path = output_path or mlconf.artifact_path
self.passthrough = passthrough
self.with_default_targets = True
@property
def entities(self) -> list[Entity]:
"""feature set entities (indexes)"""
return self._entities
@entities.setter
def entities(self, entities: list[Union[Entity, str]]):
if entities:
# if the entity is a string, convert it to Entity class
for i, entity in enumerate(entities):
if isinstance(entity, str):
entities[i] = Entity(entity)
elif isinstance(entity, Entity) and entity.name is None:
raise mlrun.errors.MLRunInvalidArgumentError(
"You have to provide an "
"Entity with valid name of string type"
)
elif isinstance(entity, dict) and (
"name" not in entity
or ("name" in entity and entity["name"] is None)
):
raise mlrun.errors.MLRunInvalidArgumentError(
"You have to provide an "
"Entity with valid name of string type"
)
self._entities = ObjectList.from_list(Entity, entities)
@property
def features(self) -> list[Feature]:
"""feature set features list"""
return self._features
@features.setter
def features(self, features: list[Feature]):
self._features = ObjectList.from_list(Feature, features)
@property
def targets(self) -> list[DataTargetBase]:
"""list of desired targets (material storage)"""
return self._targets
@targets.setter
def targets(self, targets: list[DataTargetBase]):
self._targets = ObjectList.from_list(DataTargetBase, targets)
@property
def engine(self) -> str:
"""feature set processing engine (storey, pandas, spark)"""
return self._engine
@engine.setter
def engine(self, engine: str):
engine_list = ["pandas", "spark", "storey"]
engine = engine if engine else "storey"
if engine not in engine_list:
raise mlrun.errors.MLRunInvalidArgumentError(
f"engine must be one of {','.join(engine_list)}"
)
self.graph.engine = "sync" if engine and engine in ["pandas", "spark"] else None
self._engine = engine
@property
def graph(self) -> RootFlowStep:
"""feature set transformation graph/DAG"""
return self._graph
@graph.setter
def graph(self, graph):
self._graph = self._verify_dict(graph, "graph", RootFlowStep)
self._graph.engine = (
"sync" if self.engine and self.engine in ["pandas", "spark"] else None
)
@property
def function(self) -> FunctionReference:
"""reference to template graph processing function"""
return self._function
@function.setter
def function(self, function):
self._function = self._verify_dict(function, "function", FunctionReference)
@property
def source(self) -> DataSource:
"""feature set data source definitions"""
return self._source
@source.setter
def source(self, source: Union[BaseSourceDriver, dict]):
if isinstance(source, dict):
kind = source.get("kind", "")
source = source_kind_to_driver[kind].from_dict(source)
self._source = source
@property
def relations(self) -> dict[str, Entity]:
"""feature set relations dict"""
return self._relations
@relations.setter
def relations(self, relations: dict[str, Entity]):
for col, ent in relations.items():
if isinstance(ent, str):
relations[col] = Entity(ent)
self._relations = ObjectDict.from_dict({"entity": Entity}, relations, "entity")
def require_processing(self):
return len(self._graph.steps) > 0
def validate_no_processing_for_passthrough(self):
if self.passthrough and self.require_processing():
raise mlrun.errors.MLRunInvalidArgumentError(
"passthrough feature set can not have graph transformations"
)
[docs]class FeatureSetStatus(ModelObj):
def __init__(
self,
state=None,
targets=None,
stats=None,
preview=None,
function_uri=None,
run_uri=None,
):
"""Feature set status object, containing the current feature-set's status.
.. warning::
This class should not be modified directly. It is managed by the parent feature-set object or using
feature-store APIs. Modifying the status manually may result in unpredictable behaviour.
:param state: object's current state
:param targets: list of the data targets used in the last ingestion operation
:param stats: feature statistics calculated in the last ingestion (if stats calculation was requested)
:param preview: preview of the feature-set contents (if preview generation was requested)
:param function_uri: function used to execute the feature-set graph
:param run_uri: last run used for ingestion
"""
self.state = state or mlrun.common.schemas.object.ObjectStatusState.CREATED
self._targets: ObjectList = None
self.targets = targets or []
self.stats = stats or {}
self.preview = preview or []
self.function_uri = function_uri
self.run_uri = run_uri
@property
def targets(self) -> list[DataTarget]:
"""list of material storage targets + their status/path"""
return self._targets
@targets.setter
def targets(self, targets: list[DataTarget]):
self._targets = ObjectList.from_list(DataTarget, targets)
def update_target(self, target: DataTarget):
self._targets.update(target)
def update_last_written_for_target(self, target_path: str, last_written: datetime):
for target in self._targets:
actual_target_path = get_target_driver(target).get_target_path()
if (
actual_target_path == target_path
or actual_target_path.rstrip("/") == target_path
):
target.last_written = last_written
def emit_policy_to_dict(policy: EmitPolicy):
# Storey expects the policy to be converted to a dictionary with specific params and won't allow extra params
# (see Storey's _dict_to_emit_policy function). This takes care of creating a dict conforming to it.
# TODO - fix Storey's handling of emit policy and parsing of dict in _dict_to_emit_policy.
struct = {"mode": policy.name()}
if hasattr(policy, "delay_in_seconds"):
struct["delay"] = getattr(policy, "delay_in_seconds")
if hasattr(policy, "max_events"):
struct["maxEvents"] = getattr(policy, "max_events")
return struct
[docs]class FeatureSet(ModelObj):
kind = mlrun.common.schemas.ObjectKind.feature_set.value
_dict_fields = ["kind", "metadata", "spec", "status"]
def __init__(
self,
name: str = None,
description: str = None,
entities: list[Union[Entity, str]] = None,
timestamp_key: str = None,
engine: str = None,
label_column: str = None,
relations: dict[str, Union[Entity, str]] = None,
passthrough: bool = None,
):
"""Feature set object, defines a set of features and their data pipeline
example::
import mlrun.feature_store as fstore
ticks = fstore.FeatureSet(
"ticks", entities=["stock"], timestamp_key="timestamp"
)
ticks.ingest(df)
:param name: name of the feature set
:param description: text description
:param entities: list of entity (index key) names or :py:class:`~mlrun.features.FeatureSet.Entity`
:param timestamp_key: timestamp column name
:param engine: name of the processing engine (storey, pandas, or spark), defaults to storey
:param label_column: name of the label column (the one holding the target (y) values)
:param relations: dictionary that indicates all the relations this feature set
have with another feature sets. The format of this dictionary is
{"my_column":Entity, ...}
:param passthrough: if true, ingest will skip offline targets, and get_offline_features will read
directly from source
"""
self._spec: FeatureSetSpec = None
self._metadata = None
self._status = None
self._api_client = None
self._run_db = None
self.spec = FeatureSetSpec(
description=description,
entities=entities,
timestamp_key=timestamp_key,
engine=engine,
label_column=label_column,
relations=relations,
passthrough=passthrough,
)
if timestamp_key in self.spec.entities.keys():
raise mlrun.errors.MLRunInvalidArgumentError(
"timestamp key can not be entity"
)
self.metadata = VersionedObjMetadata(name=name)
self.status = None
self._last_state = ""
self._aggregations = {}
self.set_targets()
@property
def spec(self) -> FeatureSetSpec:
return self._spec
@spec.setter
def spec(self, spec):
self._spec = self._verify_dict(spec, "spec", FeatureSetSpec)
@property
def metadata(self) -> VersionedObjMetadata:
return self._metadata
@metadata.setter
def metadata(self, metadata):
self._metadata = self._verify_dict(metadata, "metadata", VersionedObjMetadata)
@property
def status(self) -> FeatureSetStatus:
return self._status
@status.setter
def status(self, status):
self._status = self._verify_dict(status, "status", FeatureSetStatus)
@property
def uri(self):
"""fully qualified feature set uri"""
return get_store_uri(StorePrefix.FeatureSet, self.fullname)
@property
def fullname(self) -> str:
"""full name in the form ``{project}/{name}[:{tag}]``"""
fullname = (
f"{self._metadata.project or mlconf.default_project}/{self._metadata.name}"
)
if self._metadata.tag:
fullname += ":" + self._metadata.tag
return fullname
def _get_run_db(self):
if self._run_db:
return self._run_db
else:
return mlrun.get_run_db()
def _override_run_db(self, run_db):
self._run_db = run_db
[docs] def get_target_path(self, name=None):
"""get the url/path for an offline or specified data target"""
target = get_offline_target(self, name=name)
if not target and name:
target = get_online_target(self, name)
if target:
return target.get_path().get_absolute_path(
project_name=self.metadata.project
)
[docs] def set_targets(
self,
targets=None,
with_defaults=True,
default_final_step=None,
default_final_state=None,
):
"""set the desired target list or defaults
:param targets: list of target type names ('csv', 'nosql', ..) or target objects
CSVTarget(), ParquetTarget(), NoSqlTarget(), StreamTarget(), ..
:param with_defaults: add the default targets (as defined in the central config)
:param default_final_step: the final graph step after which we add the
target writers, used when the graph branches and
the end cant be determined automatically
:param default_final_state: *Deprecated* - use default_final_step instead
"""
if default_final_state:
warnings.warn(
"The 'default_final_state' parameter is deprecated in 1.3.0 and will be remove in 1.5.0. "
"Use 'default_final_step' instead.",
# TODO: remove in 1.5.0
FutureWarning,
)
default_final_step = default_final_step or default_final_state
if targets is not None and not isinstance(targets, list):
raise mlrun.errors.MLRunInvalidArgumentError(
"targets can only be None or a list of kinds or DataTargetBase derivatives"
)
targets = targets or []
if with_defaults:
self.spec.with_default_targets = True
targets.extend(get_default_targets(offline_only=self.spec.passthrough))
else:
self.spec.with_default_targets = False
self.spec.targets = []
self.__set_targets_add_targets_helper(targets)
if default_final_step:
self.spec.graph.final_step = default_final_step
def __set_targets_add_targets_helper(self, targets):
"""
Add the desired target list
:param targets: list of target type names ('csv', 'nosql', ..) or target objects
CSVTarget(), ParquetTarget(), NoSqlTarget(), StreamTarget(), ..
"""
validate_target_list(targets=targets)
for target in targets:
kind = target.kind if hasattr(target, "kind") else target
if kind not in TargetTypes.all():
raise mlrun.errors.MLRunInvalidArgumentError(
f"target kind is not supported, use one of: {','.join(TargetTypes.all())}"
)
if not hasattr(target, "kind"):
target = DataTargetBase(
target, name=str(target), partitioned=(target == "parquet")
)
self.spec.targets.update(target)
[docs] def validate_steps(self, namespace):
if not self.spec:
return
if not self.spec.graph:
return
for step in self.spec.graph.steps.values():
if (
step.class_name in queue_class_names
or step.class_name is None
or "." not in step.class_name
):
# we are not checking none class names or queue class names.
continue
class_object, class_name = step.get_step_class_object(namespace=namespace)
if not hasattr(class_object, "validate_args"):
continue
class_args = step.get_full_class_args(
namespace=namespace, class_object=class_object
)
if class_name.startswith("storey"):
class_object.validate_args(
**(class_args if class_args is not None else {})
)
else:
class_object.validate_args(
self, **(class_args if class_args is not None else {})
)
[docs] def purge_targets(self, target_names: list[str] = None, silent: bool = False):
"""Delete data of specific targets
:param target_names: List of names of targets to delete (default: delete all ingested targets)
:param silent: Fail silently if target doesn't exist in featureset status"""
verify_feature_set_permissions(
self, mlrun.common.schemas.AuthorizationAction.delete
)
purge_targets = self._reload_and_get_status_targets(
target_names=target_names, silent=silent
)
if purge_targets:
purge_target_names = list(purge_targets.keys())
for target_name in purge_target_names:
target = purge_targets[target_name]
driver = get_target_driver(target_spec=target, resource=self)
try:
driver.purge()
except FileNotFoundError:
pass
del self.status.targets[target_name]
self.save()
[docs] def update_targets_for_ingest(
self,
targets: list[DataTargetBase],
overwrite: bool = None,
):
if not targets:
return
ingestion_target_names = [t.name for t in targets]
status_targets = {}
if not overwrite:
# silent=True always because targets are not guaranteed to be found in status
status_targets = (
self._reload_and_get_status_targets(
target_names=ingestion_target_names, silent=True
)
or {}
)
update_targets_run_id_for_ingest(overwrite, targets, status_targets)
def _reload_and_get_status_targets(
self, target_names: list[str] = None, silent: bool = False
):
try:
self.reload(update_spec=False)
except mlrun.errors.MLRunNotFoundError:
# If the feature set doesn't exist in DB there shouldn't be any target to delete
if silent:
return
else:
raise
if target_names:
targets = ObjectList(DataTarget)
for target_name in target_names:
try:
targets[target_name] = self.status.targets[target_name]
except KeyError:
if silent:
pass
else:
raise mlrun.errors.MLRunNotFoundError(
f"Target not found in status (fset={self.metadata.name}, target={target_name})"
)
else:
targets = self.status.targets
return targets
[docs] def has_valid_source(self):
"""check if object's spec has a valid (non empty) source definition"""
source = self.spec.source
return source is not None and source.path is not None and source.path != "None"
[docs] def add_entity(
self,
name: str,
value_type: mlrun.data_types.ValueType = None,
description: str = None,
labels: Optional[dict[str, str]] = None,
):
"""add/set an entity (dataset index)
example::
import mlrun.feature_store as fstore
ticks = fstore.FeatureSet(
"ticks", entities=["stock"], timestamp_key="timestamp"
)
ticks.add_entity(
"country", mlrun.data_types.ValueType.STRING, description="stock country"
)
ticks.add_entity("year", mlrun.data_types.ValueType.INT16)
ticks.save()
:param name: entity name
:param value_type: type of the entity (default to ValueType.STRING)
:param description: description of the entity
:param labels: label tags dict
"""
entity = Entity(name, value_type, description=description, labels=labels)
self._spec.entities.update(entity, name)
[docs] def add_feature(self, feature: mlrun.features.Feature, name=None):
"""add/set a feature
example::
import mlrun.feature_store as fstore
from mlrun.features import Feature
ticks = fstore.FeatureSet(
"ticks", entities=["stock"], timestamp_key="timestamp"
)
ticks.add_feature(
Feature(
value_type=mlrun.data_types.ValueType.STRING,
description="client consistency",
),
"ABC01",
)
ticks.add_feature(
Feature(
value_type=mlrun.data_types.ValueType.FLOAT,
description="client volatility",
),
"SAB",
)
ticks.save()
:param feature: setting of Feature
:param name: feature name
"""
self._spec.features.update(feature, name)
[docs] def link_analysis(self, name, uri):
"""add a linked file/artifact (chart, data, ..)"""
self._spec.analysis[name] = uri
@property
def graph(self) -> RootFlowStep:
"""feature set transformation graph/DAG"""
return self.spec.graph
def _add_aggregation_to_existing(self, new_aggregation):
name = new_aggregation["name"]
if name in self._aggregations:
current_aggr = self._aggregations[name]
if current_aggr["windows"] != new_aggregation["windows"]:
raise mlrun.errors.MLRunInvalidArgumentError(
f"Aggregation with name {name} already exists but with window {current_aggr['windows']}. "
f"Please provide name for the aggregation"
)
if current_aggr["period"] != new_aggregation["period"]:
raise mlrun.errors.MLRunInvalidArgumentError(
f"Aggregation with name {name} already exists but with period {current_aggr['period']}. "
f"Please provide name for the aggregation"
)
if current_aggr["column"] != new_aggregation["column"]:
raise mlrun.errors.MLRunInvalidArgumentError(
f"Aggregation with name {name} already exists but for different column {current_aggr['column']}. "
f"Please provide name for the aggregation"
)
current_aggr["operations"] = list(
set(current_aggr["operations"] + new_aggregation["operations"])
)
return
self._aggregations[name] = new_aggregation
[docs] def add_aggregation(
self,
column,
operations,
windows,
period=None,
name=None,
step_name=None,
after=None,
before=None,
emit_policy: EmitPolicy = None,
):
"""add feature aggregation rule
example::
myset.add_aggregation("ask", ["sum", "max"], "1h", "10m", name="asks")
:param column: name of column/field aggregate. Do not name columns starting with either `_` or `aggr_`.
They are reserved for internal use, and the data does not ingest correctly.
When using the pandas engine, do not use spaces (` `) or periods (`.`) in the column names;
they cause errors in the ingestion.
:param operations: aggregation operations. Supported operations:
count, sum, sqr, max, min, first, last, avg, stdvar, stddev
:param windows: time windows, can be a single window, e.g. '1h', '1d',
or a list of same unit windows e.g. ['1h', '6h']
windows are transformed to fixed windows or
sliding windows depending whether period parameter
provided.
- Sliding window is fixed-size overlapping windows
that slides with time.
The window size determines the size of the sliding window
and the period determines the step size to slide.
Period must be integral divisor of the window size.
If the period is not provided then fixed windows is used.
- Fixed window is fixed-size, non-overlapping, gap-less window.
The window is referred to as a tumbling window.
In this case, each record on an in-application stream belongs
to a specific window. It is processed only once
(when the query processes the window to which the record belongs).
:param period: optional, sliding window granularity, e.g. '20s' '10m' '3h' '7d'
:param name: optional, aggregation name/prefix. Must be unique per feature set. If not passed,
the column will be used as name.
:param step_name: optional, graph step name
:param after: optional, after which graph step it runs
:param before: optional, comes before graph step
:param emit_policy: optional, which emit policy to use when performing the aggregations. Use the derived
classes of ``storey.EmitPolicy``. The default is to emit every period for Spark engine
and emit every event for storey. Currently the only other supported option is to use
``emit_policy=storey.EmitEveryEvent()`` when using the Spark engine to emit every event
"""
if isinstance(operations, str):
raise mlrun.errors.MLRunInvalidArgumentError(
"Invalid parameters provided - operations must be a list."
)
name = name or column
if isinstance(windows, str):
windows = [windows]
if isinstance(operations, str):
operations = [operations]
aggregation = FeatureAggregation(
name, column, operations, windows, period
).to_dict()
def upsert_feature(name):
if name in self.spec.features:
self.spec.features[name].aggregate = True
else:
self.spec.features[name] = Feature(
name=column, aggregate=True, value_type="float"
)
step_name = step_name or aggregates_step
graph = self.spec.graph
if step_name in graph.steps:
step = graph.steps[step_name]
self._add_aggregation_to_existing(aggregation)
step.class_args["aggregates"] = list(self._aggregations.values())
if emit_policy and self.spec.engine == "spark":
# Using simple override here - we might want to consider exploding if different emit policies
# were used for multiple aggregations.
emit_policy_dict = emit_policy_to_dict(emit_policy)
if "emit_policy" in step.class_args:
curr_emit_policy = step.class_args["emit_policy"]["mode"]
if curr_emit_policy != emit_policy_dict["mode"]:
logger.warning(
f"Current emit policy will be overridden: {curr_emit_policy} => {emit_policy_dict['mode']}"
)
step.class_args["emit_policy"] = emit_policy_dict
else:
class_args = {}
self._aggregations[aggregation["name"]] = aggregation
if before is None and after is None:
after = previous_step
if not self.spec.engine or self.spec.engine == "storey":
step = graph.add_step(
name=step_name,
after=after,
before=before,
class_name="storey.AggregateByKey",
time_field=self.spec.timestamp_key,
aggregates=[aggregation],
table=".",
**class_args,
)
elif self.spec.engine == "spark":
key_columns = []
if emit_policy:
class_args["emit_policy"] = emit_policy_to_dict(emit_policy)
for entity in self.spec.entities:
key_columns.append(entity.name)
step = graph.add_step(
name=step_name,
key_columns=key_columns,
time_column=self.spec.timestamp_key,
aggregates=[aggregation],
after=after,
before=before,
class_name="mlrun.feature_store.feature_set.SparkAggregateByKey",
**class_args,
)
else:
raise ValueError(
"Aggregations are only implemented for storey and spark engines."
)
for operation in operations:
for window in windows:
upsert_feature(f"{name}_{operation}_{window}")
return step
[docs] def get_stats_table(self):
"""get feature statistics table (as dataframe)"""
if self.status.stats:
return pd.DataFrame.from_dict(self.status.stats, orient="index")
def __getitem__(self, name):
return self._spec.features[name]
def __setitem__(self, key, item):
if key not in self._spec.entities.keys():
self._spec.features.update(item, key)
else:
raise mlrun.errors.MLRunInvalidArgumentError(
"A `FeatureSet` cannot have an entity and a feature with the same name. "
f"The feature that was given to add '{key}' has the same name of the `FeatureSet`'s entity."
)
[docs] def plot(self, filename=None, format=None, with_targets=False, **kw):
"""plot/save graph using graphviz
example::
import mlrun.feature_store as fstore
...
ticks = fstore.FeatureSet(
"ticks", entities=["stock"], timestamp_key="timestamp"
)
ticks.add_aggregation(
name="priceN",
column="price",
operations=["avg"],
windows=["1d"],
period="1h",
)
ticks.plot(rankdir="LR", with_targets=True)
:param filename: target filepath for the graph image (None for the notebook)
:param format: the output format used for rendering (``'pdf'``, ``'png'``, etc.)
:param with_targets: show targets in the graph image
:param kw: kwargs passed to graphviz, e.g. rankdir=”LR” (see https://graphviz.org/doc/info/attrs.html)
:return: graphviz graph object
"""
graph = self.spec.graph
_, default_final_step, _ = graph.check_and_process_graph(allow_empty=True)
targets = None
if with_targets:
validate_target_list(targets=targets)
validate_target_placement(graph, default_final_step, self.spec.targets)
targets = [
BaseStep(
f"{target.kind}/{target.name}",
after=target.after_step or default_final_step,
shape="cylinder",
)
for target in self.spec.targets
]
return graph.plot(filename, format, targets=targets, **kw)
[docs] def to_dataframe(
self,
columns=None,
df_module=None,
target_name=None,
start_time=None,
end_time=None,
time_column=None,
additional_filters=None,
**kwargs,
):
"""return featureset (offline) data as dataframe
:param columns: list of columns to select (if not all)
:param df_module: py module used to create the DataFrame (pd for Pandas, dd for Dask, ..)
:param target_name: select a specific target (material view)
:param start_time: filter by start time
:param end_time: filter by end time
:param time_column: specify the time column name in the file
:param kwargs: additional reader (csv, parquet, ..) args
:param additional_filters: List of additional_filter conditions as tuples.
Each tuple should be in the format (column_name, operator, value).
Supported operators: "=", ">=", "<=", ">", "<".
Example: [("Product", "=", "Computer")]
For all supported filters, please see:
https://arrow.apache.org/docs/python/generated/pyarrow.parquet.ParquetDataset.html
:return: DataFrame
"""
entities = list(self.spec.entities.keys())
if columns:
if self.spec.timestamp_key and self.spec.timestamp_key not in entities:
columns = [self.spec.timestamp_key] + columns
columns = entities + columns
if self.spec.passthrough:
if not self.spec.source:
raise mlrun.errors.MLRunNotFoundError(
"passthrough feature set {self.metadata.name} with no source"
)
df = self.spec.source.to_dataframe(
columns=columns,
start_time=start_time,
end_time=end_time,
time_field=time_column,
additional_filters=additional_filters,
**kwargs,
)
# to_dataframe() can sometimes return an iterator of dataframes instead of one dataframe
if not isinstance(df, pd.DataFrame):
df = pd.concat(df)
return df
target = get_offline_target(self, name=target_name)
if not target:
raise mlrun.errors.MLRunNotFoundError(
"there are no offline targets for this feature set"
)
result = target.as_df(
columns=columns,
df_module=df_module,
entities=entities,
start_time=start_time,
end_time=end_time,
time_column=time_column,
additional_filters=additional_filters,
**kwargs,
)
return result
[docs] def save(self, tag="", versioned=False):
"""save to mlrun db"""
db = self._get_run_db()
self.metadata.project = self.metadata.project or mlconf.default_project
tag = tag or self.metadata.tag or "latest"
as_dict = self.to_dict()
as_dict["spec"]["features"] = as_dict["spec"].get(
"features", []
) # bypass DB bug
db.store_feature_set(as_dict, tag=tag, versioned=versioned)
[docs] def reload(self, update_spec=True):
"""reload/sync the feature vector status and spec from the DB"""
feature_set = self._get_run_db().get_feature_set(
self.metadata.name, self.metadata.project, self.metadata.tag
)
if isinstance(feature_set, dict):
feature_set = FeatureSet.from_dict(feature_set)
self.status = feature_set.status
if update_spec:
self.spec = feature_set.spec
[docs] def ingest(
self,
source=None,
targets: list[DataTargetBase] = None,
namespace=None,
return_df: bool = True,
infer_options: InferOptions = InferOptions.default(),
run_config: RunConfig = None,
mlrun_context=None,
spark_context=None,
overwrite=None,
) -> Optional[pd.DataFrame]:
"""Read local DataFrame, file, URL, or source into the feature store
Ingest reads from the source, run the graph transformations, infers metadata and stats
and writes the results to the default of specified targets
when targets are not specified data is stored in the configured default targets
(will usually be NoSQL for real-time and Parquet for offline).
the `run_config` parameter allow specifying the function and job configuration,
see: :py:class:`~mlrun.feature_store.RunConfig`
example::
stocks_set = FeatureSet("stocks", entities=[Entity("ticker")])
stocks = pd.read_csv("stocks.csv")
df = stocks_set.ingest(stocks, infer_options=fstore.InferOptions.default())
# for running as remote job
config = RunConfig(image="mlrun/mlrun")
df = ingest(stocks_set, stocks, run_config=config)
# specify source and targets
source = CSVSource("mycsv", path="measurements.csv")
targets = [CSVTarget("mycsv", path="./mycsv.csv")]
ingest(measurements, source, targets)
:param source: source dataframe or other sources (e.g. parquet source see:
:py:class:`~mlrun.datastore.ParquetSource` and other classes in mlrun.datastore with suffix
Source)
:param targets: optional list of data target objects
:param namespace: namespace or module containing graph classes
:param return_df: indicate if to return a dataframe with the graph results
:param infer_options: schema (for discovery of entities, features in featureset), index, stats,
histogram and preview infer options (:py:class:`~mlrun.feature_store.InferOptions`)
:param run_config: function and/or run configuration for remote jobs,
see :py:class:`~mlrun.feature_store.RunConfig`
:param mlrun_context: mlrun context (when running as a job), for internal use !
:param spark_context: local spark session for spark ingestion, example for creating the spark context:
`spark = SparkSession.builder.appName("Spark function").getOrCreate()`
For remote spark ingestion, this should contain the remote spark service name
:param overwrite: delete the targets' data prior to ingestion
(default: True for non scheduled ingest - deletes the targets that are about to be ingested.
False for scheduled ingest - does not delete the target)
:return: if return_df is True, a dataframe will be returned based on the graph
"""
return mlrun.feature_store.api._ingest(
self,
source,
targets,
namespace,
return_df,
infer_options,
run_config,
mlrun_context,
spark_context,
overwrite,
)
[docs] def preview(
self,
source,
entity_columns: list = None,
namespace=None,
options: InferOptions = None,
verbose: bool = False,
sample_size: int = None,
) -> pd.DataFrame:
"""run the ingestion pipeline with local DataFrame/file data and infer features schema and stats
example::
quotes_set = FeatureSet("stock-quotes", entities=[Entity("ticker")])
quotes_set.add_aggregation("ask", ["sum", "max"], ["1h", "5h"], "10m")
quotes_set.add_aggregation("bid", ["min", "max"], ["1h"], "10m")
df = quotes_set.preview(
quotes_df,
entity_columns=["ticker"],
)
:param source: source dataframe or csv/parquet file path
:param entity_columns: list of entity (index) column names
:param namespace: namespace or module containing graph classes
:param options: schema (for discovery of entities, features in featureset), index, stats,
histogram and preview infer options (:py:class:`~mlrun.feature_store.InferOptions`)
:param verbose: verbose log
:param sample_size: num of rows to sample from the dataset (for large datasets)
"""
return mlrun.feature_store.api._preview(
self, source, entity_columns, namespace, options, verbose, sample_size
)
[docs] def deploy_ingestion_service(
self,
source: DataSource = None,
targets: list[DataTargetBase] = None,
name: str = None,
run_config: RunConfig = None,
verbose=False,
) -> tuple[str, BaseRuntime]:
"""Start real-time ingestion service using nuclio function
Deploy a real-time function implementing feature ingestion pipeline
the source maps to Nuclio event triggers (http, kafka, v3io stream, etc.)
the `run_config` parameter allow specifying the function and job configuration,
see: :py:class:`~mlrun.feature_store.RunConfig`
example::
source = HTTPSource()
func = mlrun.code_to_function("ingest", kind="serving").apply(mount_v3io())
config = RunConfig(function=func)
my_set.deploy_ingestion_service(source, run_config=config)
:param source: data source object describing the online or offline source
:param targets: list of data target objects
:param name: name for the job/function
:param run_config: service runtime configuration (function object/uri, resources, etc..)
:param verbose: verbose log
:return: URL to access the deployed ingestion service, and the function that was deployed (which will
differ from the function passed in via the run_config parameter).
"""
return mlrun.feature_store.api._deploy_ingestion_service_v2(
self, source, targets, name, run_config, verbose
)
[docs] def is_connectable_to_df(self, df_columns: list[str]) -> bool:
"""
This method checks if the dataframe can be left-joined with this feature set
:param df_columns: The columns of the data frame you want to merge to the left of this feature set
:return: True if it can be left-joined and False otherwise
"""
return df_columns and all(
ent in df_columns for ent in self.spec.entities.keys()
)
class SparkAggregateByKey(StepToDict):
_supported_operations = [
"count",
"sum",
"sqr",
"max",
"min",
"first",
"last",
"avg",
"stdvar",
"stddev",
]
def __init__(
self,
key_columns: list[str],
time_column: str,
aggregates: list[dict],
emit_policy: Union[EmitPolicy, dict] = None,
):
self.key_columns = key_columns
self.time_column = time_column
self.aggregates = aggregates
self.emit_policy_mode = None
if emit_policy:
if isinstance(emit_policy, EmitPolicy):
emit_policy = emit_policy_to_dict(emit_policy)
self.emit_policy_mode = emit_policy["mode"]
@staticmethod
def _duration_to_spark_format(duration):
num = duration[:-1]
unit = duration[-1:]
if unit == "d":
unit = "day"
elif unit == "h":
unit = "hour"
elif unit == "m":
unit = "minute"
elif unit == "s":
unit = "second"
else:
raise ValueError(f"Invalid duration '{duration}'")
return f"{num} {unit}"
@staticmethod
def _verify_operation(op):
if op not in SparkAggregateByKey._supported_operations:
error_string = (
f"operation {op} is unsupported. Supported operations: "
+ ", ".join(SparkAggregateByKey._supported_operations)
)
raise mlrun.errors.MLRunInvalidArgumentError(error_string)
def _extract_fields_from_aggregate_dict(self, aggregate):
name = aggregate["name"]
column = aggregate["column"]
operations = aggregate["operations"]
for op in operations:
self._verify_operation(op)
windows = aggregate["windows"]
spark_period = (
self._duration_to_spark_format(aggregate["period"])
if "period" in aggregate
else None
)
return name, column, operations, windows, spark_period
@staticmethod
def _get_aggr(operation, column):
import pyspark.sql.functions as funcs
if operation == "sqr":
return funcs.sum(funcs.expr(f"{column} * {column}"))
elif operation == "stdvar":
return funcs.variance(column)
else:
func = getattr(funcs, operation)
return func(column)
def do(self, event):
import pyspark.sql.functions as funcs
from pyspark.sql import Window
time_column = self.time_column or "time"
input_df = event
if not self.emit_policy_mode or self.emit_policy_mode != EmitEveryEvent.name():
last_value_aggs = [
funcs.last(column).alias(column)
for column in input_df.columns
if column not in self.key_columns and column != time_column
]
dfs = []
for aggregate in self.aggregates:
(
name,
column,
operations,
windows,
spark_period,
) = self._extract_fields_from_aggregate_dict(aggregate)
for window in windows:
spark_window = self._duration_to_spark_format(window)
aggs = last_value_aggs
for operation in operations:
agg = self._get_aggr(operation, column)
agg_name = f"{name if name else column}_{operation}_{window}"
agg = agg.alias(agg_name)
aggs.append(agg)
window_column = funcs.window(
time_column, spark_window, spark_period
)
df = input_df.groupBy(
*self.key_columns,
window_column.end.alias(time_column),
).agg(*aggs)
df = df.withColumn(f"{time_column}_window", funcs.lit(window))
dfs.append(df)
union_df = dfs[0]
for df in dfs[1:]:
union_df = union_df.unionByName(df, allowMissingColumns=True)
return union_df
else:
window_counter = 0
# We'll use this column to identify our original row and group-by across the various windows
# (either sliding windows or multiple windows provided). See below comment for more details.
rowid_col = "__mlrun_rowid"
df = input_df.withColumn(rowid_col, funcs.monotonically_increasing_id())
drop_columns = [rowid_col]
window_rank_cols = []
union_df = None
for aggregate in self.aggregates:
(
name,
column,
operations,
windows,
spark_period,
) = self._extract_fields_from_aggregate_dict(aggregate)
for window in windows:
spark_window = self._duration_to_spark_format(window)
window_col = f"__mlrun_window_{window_counter}"
win_df = df.withColumn(
window_col,
funcs.window(time_column, spark_window, spark_period).end,
)
function_window = Window.partitionBy(*self.key_columns, window_col)
window_rank_col = f"__mlrun_win_rank_{window_counter}"
rank_window = Window.partitionBy(rowid_col).orderBy(window_col)
win_df = win_df.withColumn(
window_rank_col, funcs.row_number().over(rank_window)
)
window_rank_cols.append(window_rank_col)
drop_columns.extend([window_col, window_rank_col])
window_counter += 1
for operation in operations:
agg = self._get_aggr(operation, column)
agg_name = f"{name if name else column}_{operation}_{window}"
win_df = win_df.withColumn(agg_name, agg.over(function_window))
union_df = (
union_df.unionByName(win_df, allowMissingColumns=True)
if union_df
else win_df
)
# We need to collapse the multiple window rows that were generated during the query processing. For that
# purpose we'll pick just the 1st row for each window, and then group-by with ignorenulls. Basically since
# the result is a union of multiple windows, we'll get something like this for each input row:
# row window_1 rank_window_1 window_2 rank_window_2 ...calculations and fields...
# 1 10:00 1 null null ...
# 2 10:10 2 null null ...
# 3 null null 10:00 1 ...
# 4 null null 10:10 2 ...
# And we want to take rows 1 and 3 in this case. Then the group-by will merge them to a single line since
# it ignores nulls, so it will take the values for window_1 from row 1 and for window_2 from row 3.
window_filter = " or ".join(
[f"{window_rank_col} == 1" for window_rank_col in window_rank_cols]
)
first_value_aggs = [
funcs.first(column, ignorenulls=True).alias(column)
for column in union_df.columns
if column not in drop_columns
]
return (
union_df.filter(window_filter)
.groupBy(rowid_col)
.agg(*first_value_aggs)
.drop(rowid_col)
)
