계보 엔터티 쿼리

from sagemaker.lineage.context import Context, EndpointContext
from sagemaker.lineage.action import Action
from sagemaker.lineage.association import Association
from sagemaker.lineage.artifact import Artifact, ModelArtifact, DatasetArtifact

from sagemaker.lineage.query import (
    LineageQuery,
    LineageFilter,
    LineageSourceEnum,
    LineageEntityEnum,
    LineageQueryDirectionEnum,
)
# Find the endpoint context and model artifact that should be used for the lineage queries.

contexts = Context.list(source_uri=endpoint_arn)
context_name = list(contexts)[0].context_name
endpoint_context = EndpointContext.load(context_name=context_name)
        

# Define the LineageFilter to look for entities of type `ARTIFACT` and the source of type `DATASET`.

query_filter = LineageFilter(
    entities=[LineageEntityEnum.ARTIFACT], sources=[LineageSourceEnum.DATASET]
)

# Providing this `LineageFilter` to the `LineageQuery` constructs a query that traverses through the given context `endpoint_context`
# and find all datasets.

query_result = LineageQuery(sagemaker_session).query(
    start_arns=[endpoint_context.context_arn],
    query_filter=query_filter,
    direction=LineageQueryDirectionEnum.ASCENDANTS,
    include_edges=False,
)

# Parse through the query results to get the lineage objects corresponding to the datasets
dataset_artifacts = []
for vertex in query_result.vertices:
    dataset_artifacts.append(vertex.to_lineage_object().source.source_uri)

pp.pprint(dataset_artifacts)

# Define the LineageFilter to look for entities of type `ARTIFACT` and the source of type `MODEL`.

query_filter = LineageFilter(
    entities=[LineageEntityEnum.ARTIFACT], sources=[LineageSourceEnum.MODEL]
)

# Providing this `LineageFilter` to the `LineageQuery` constructs a query that traverses through the given context `endpoint_context`
# and find all datasets.

query_result = LineageQuery(sagemaker_session).query(
    start_arns=[endpoint_context.context_arn],
    query_filter=query_filter,
    direction=LineageQueryDirectionEnum.ASCENDANTS,
    include_edges=False,
)

# Parse through the query results to get the lineage objects corresponding to the model
model_artifacts = []
for vertex in query_result.vertices:
    model_artifacts.append(vertex.to_lineage_object().source.source_uri)

# The results of the `LineageQuery` API call return the ARN of the model deployed to the endpoint along with
# the S3 URI to the model.tar.gz file associated with the model
pp.pprint(model_artifacts)

# Define the LineageFilter to look for entities of type `TRIAL_COMPONENT` and the source of type `TRAINING_JOB`.

query_filter = LineageFilter(
    entities=[LineageEntityEnum.TRIAL_COMPONENT],
    sources=[LineageSourceEnum.TRAINING_JOB],
)

# Providing this `LineageFilter` to the `LineageQuery` constructs a query that traverses through the given context `endpoint_context`
# and find all datasets.

query_result = LineageQuery(sagemaker_session).query(
    start_arns=[endpoint_context.context_arn],
    query_filter=query_filter,
    direction=LineageQueryDirectionEnum.ASCENDANTS,
    include_edges=False,
)

# Parse through the query results to get the ARNs of the training jobs associated with this Endpoint
trial_components = []
for vertex in query_result.vertices:
    trial_components.append(vertex.arn)

pp.pprint(trial_components)

계보의 초점이 바뀌는 다른 start_arns을 포함하도록 LineageQuery를 수정할 수 있습니다. 또한 LineageFilter는 여러 소스와 엔터티를 사용하여 쿼리 범위를 확장할 수 있습니다.

다음에서는 모델을 계보 초점으로 사용하여 관련 엔드포인트와 데이터 세트를 찾습니다.

# Get the ModelArtifact

model_artifact_summary = list(Artifact.list(source_uri=model_package_arn))[0]
model_artifact = ModelArtifact.load(artifact_arn=model_artifact_summary.artifact_arn)
query_filter = LineageFilter(
    entities=[LineageEntityEnum.ARTIFACT],
    sources=[LineageSourceEnum.ENDPOINT, LineageSourceEnum.DATASET],
)

query_result = LineageQuery(sagemaker_session).query(
    start_arns=[model_artifact.artifact_arn],  # Model is the starting artifact
    query_filter=query_filter,
    # Find all the entities that descend from the model, i.e. the endpoint
    direction=LineageQueryDirectionEnum.DESCENDANTS,
    include_edges=False,
)

associations = []
for vertex in query_result.vertices:
    associations.append(vertex.to_lineage_object().source.source_uri)

query_result = LineageQuery(sagemaker_session).query(
    start_arns=[model_artifact.artifact_arn],  # Model is the starting artifact
    query_filter=query_filter,
    # Find all the entities that ascend from the model, i.e. the datasets
    direction=LineageQueryDirectionEnum.ASCENDANTS,
    include_edges=False,
)

for vertex in query_result.vertices:
    associations.append(vertex.to_lineage_object().source.source_uri)

pp.pprint(associations)

방향이 BOTH로 설정된 경우 쿼리는 그래프를 탐색하여 상위 및 하위 관계를 찾습니다. 이 순회는 시작 노드뿐만 아니라 방문한 각 노드에서도 발생합니다. 예를 들어, 훈련 작업을 두 번 실행하고 훈련 작업에서 생성된 두 모델을 모두 엔드포인트에 배포한 경우 방향이 BOTH로 설정된 쿼리 결과에 두 엔드포인트가 모두 표시됩니다. 이는 동일한 이미지가 모델 훈련 및 배포에 사용되기 때문입니다. 이미지는 모델에 공통적이므로, start_arn과 양쪽 엔드포인트가 쿼리 결과에 나타납니다.

query_filter = LineageFilter(
    entities=[LineageEntityEnum.ARTIFACT],
    sources=[LineageSourceEnum.ENDPOINT, LineageSourceEnum.DATASET],
)

query_result = LineageQuery(sagemaker_session).query(
    start_arns=[model_artifact.artifact_arn],  # Model is the starting artifact
    query_filter=query_filter,
    # This specifies that the query should look for associations both ascending and descending for the start
    direction=LineageQueryDirectionEnum.BOTH,
    include_edges=False,
)

associations = []
for vertex in query_result.vertices:
    associations.append(vertex.to_lineage_object().source.source_uri)

pp.pprint(associations)

계보 그래프의 방향을 이해하려면 다음 엔터티 관계 그래프(데이터 세트 -> 훈련 작업 -> 모델 -> 엔드포인트)를 참고하세요.

엔드포인트는 모델의 하위이고, 모델은 데이터 세트의 하위입니다. 마찬가지로, 모델은 엔드포인트의 상위입니다. direction 파라미터를 사용하여 start_arns에서 엔터티의 하위 항목이나 상위 항목인 엔터티를 반환할지 여부를 지정할 수 있습니다. start_arns에 모델이 포함되어 있고 방향이 DESCENDANTS인 경우 쿼리는 엔드포인트를 반환합니다. 방향이 ASCENDANTS인 경우 쿼리는 데이터 세트를 반환합니다.

# In this example, we'll look at the impact of specifying the direction as ASCENDANT or DESCENDANT in a `LineageQuery`.

query_filter = LineageFilter(
    entities=[LineageEntityEnum.ARTIFACT],
    sources=[
        LineageSourceEnum.ENDPOINT,
        LineageSourceEnum.MODEL,
        LineageSourceEnum.DATASET,
        LineageSourceEnum.TRAINING_JOB,
    ],
)

query_result = LineageQuery(sagemaker_session).query(
    start_arns=[model_artifact.artifact_arn],
    query_filter=query_filter,
    direction=LineageQueryDirectionEnum.ASCENDANTS,
    include_edges=False,
)

ascendant_artifacts = []

# The lineage entity returned for the Training Job is a TrialComponent which can't be converted to a
# lineage object using the method `to_lineage_object()` so we extract the TrialComponent ARN.
for vertex in query_result.vertices:
    try:
        ascendant_artifacts.append(vertex.to_lineage_object().source.source_uri)
    except:
        ascendant_artifacts.append(vertex.arn)

print("Ascendant artifacts : ")
pp.pprint(ascendant_artifacts)

query_result = LineageQuery(sagemaker_session).query(
    start_arns=[model_artifact.artifact_arn],
    query_filter=query_filter,
    direction=LineageQueryDirectionEnum.DESCENDANTS,
    include_edges=False,
)

descendant_artifacts = []
for vertex in query_result.vertices:
    try:
        descendant_artifacts.append(vertex.to_lineage_object().source.source_uri)
    except:
        # Handling TrialComponents.
        descendant_artifacts.append(vertex.arn)

print("Descendant artifacts : ")
pp.pprint(descendant_artifacts)

클래스 EndpointContext, ModelArtifact, 에는 특정 계보 쿼리를 더 쉽게 활용할 수 있도록 LineageQuery API 를 통해 래퍼를 지정하는 헬퍼 함수가 DatasetArtifact 있습니다. 다음 예시에서는 이러한 도우미 함수를 사용하는 방법을 보여 줍니다.

# Find all the datasets associated with this endpoint

datasets = []
dataset_artifacts = endpoint_context.dataset_artifacts()
for dataset in dataset_artifacts:
    datasets.append(dataset.source.source_uri)
print("Datasets : ", datasets)

# Find the training jobs associated with the endpoint
training_job_artifacts = endpoint_context.training_job_arns()
training_jobs = []
for training_job in training_job_artifacts:
    training_jobs.append(training_job)
print("Training Jobs : ", training_jobs)

# Get the ARN for the pipeline execution associated with this endpoint (if any)
pipeline_executions = endpoint_context.pipeline_execution_arn()
if pipeline_executions:
    for pipeline in pipelines_executions:
        print(pipeline)

# Here we use the `ModelArtifact` class to find all the datasets and endpoints associated with the model

dataset_artifacts = model_artifact.dataset_artifacts()
endpoint_contexts = model_artifact.endpoint_contexts()

datasets = [dataset.source.source_uri for dataset in dataset_artifacts]
endpoints = [endpoint.source.source_uri for endpoint in endpoint_contexts]

print("Datasets associated with this model : ")
pp.pprint(datasets)

print("Endpoints associated with this model : ")
pp.pprint(endpoints)

# Here we use the `DatasetArtifact` class to find all the endpoints hosting models that were trained with a particular dataset
# Find the artifact associated with the dataset

dataset_artifact_arn = list(Artifact.list(source_uri=training_data))[0].artifact_arn
dataset_artifact = DatasetArtifact.load(artifact_arn=dataset_artifact_arn)

# Find the endpoints that used this training dataset
endpoint_contexts = dataset_artifact.endpoint_contexts()
endpoints = [endpoint.source.source_uri for endpoint in endpoint_contexts]

print("Endpoints associated with the training dataset {}".format(training_data))
pp.pprint(endpoints)

샘플 노트북 visualizer.py에는 계보 그래프를 그리는 데 도움이 되는 도우미 클래스 Visualizer가 제공됩니다. 쿼리 응답이 렌더링되면 StartArns의 계보 관계가 표시된 그래프가 표시됩니다. 시각화StartArns에서 query_lineage API 작업에 반환된 다른 계보 엔터티와의 관계를 보여줍니다.

# Graph APIs
# Here we use the boto3 `query_lineage` API to generate the query response to plot.

from visualizer import Visualizer

query_response = sm_client.query_lineage(
    StartArns=[endpoint_context.context_arn], Direction="Ascendants", IncludeEdges=True
)

viz = Visualizer()
viz.render(query_response, "Endpoint")
        
        query_response = sm_client.query_lineage(
    StartArns=[model_artifact.artifact_arn], Direction="Ascendants", IncludeEdges=True
)
viz.render(query_response, "Model")