def k8s_sklearn_model(sklearn_model: Model, k8s_runtime: SeldonKubernetesRuntime) -> Generator[Model, None, None]:
sklearn_model.runtime = k8s_runtime
sklearn_model.deploy()
sklearn_model.wait_ready(timeout_secs=60)
yield sklearn_model
sklearn_model.undeploy()
def k8s_inference_pipeline(
sklearn_model: Model,
xgboost_model: Model,
k8s_runtime: SeldonKubernetesRuntime,
k8s_runtime_v2: SeldonKubernetesRuntime,
):
sklearn_model.runtime = k8s_runtime
xgboost_model.runtime = k8s_runtime
@pipeline(
name="classifier",
runtime=k8s_runtime_v2,
models=[sklearn_model, xgboost_model],
uri="gs://seldon-models/tempo/test",
)
def _pipeline(payload: np.ndarray) -> np.ndarray:
res1 = sklearn_model(payload=payload)
if res1[0][0] > 0.7:
return res1
else:
return xgboost_model(payload=payload)
_pipeline.save()
_pipeline.upload()
_pipeline.deploy()
_pipeline.wait_ready(timeout_secs=60)
# TODO: Pipeline shouldn't become ready until models are ready
sklearn_model.wait_ready(timeout_secs=60)
xgboost_model.wait_ready(timeout_secs=60)
yield _pipeline
try:
_pipeline.undeploy()
except docker.errors.NotFound:
# Ignore if the pipeline was already undeployed
pass