Amazon Bedrock - Model Customization Jobs

airflow/providers/amazon/aws/hooks/bedrock.py

@@ -16,9 +16,53 @@
 # under the License.
 from __future__ import annotations
 
+from botocore.exceptions import ClientError
+
 from airflow.providers.amazon.aws.hooks.base_aws import AwsBaseHook
 
 
+class BedrockHook(AwsBaseHook):
+    """
+    Interact with Amazon Bedrock.
+
+    Provide thin wrapper around :external+boto3:py:class:`boto3.client("bedrock") <Bedrock.Client>`.
+
+    Additional arguments (such as ``aws_conn_id``) may be specified and
+    are passed down to the underlying AwsBaseHook.
+
+    .. seealso::
+        - :class:`airflow.providers.amazon.aws.hooks.base_aws.AwsBaseHook`
+    """
+
+    client_type = "bedrock"
+
+    def __init__(self, *args, **kwargs) -> None:
+        kwargs["client_type"] = self.client_type
+        super().__init__(*args, **kwargs)
+
+    def _get_job_by_name(self, job_name: str):
+        return self.conn.get_model_customization_job(jobIdentifier=job_name)
+
+    def get_customize_model_job_state(self, job_name) -> str:
+        state = self._get_job_by_name(job_name)["status"]
+        self.log.info("Job '%s' state: %s", job_name, state)
+        return state
+
+    def job_name_exists(self, job_name: str) -> bool:
+        try:
+            self._get_job_by_name(job_name)
+            self.log.info("Verified that job name '%s' does exist.", job_name)
+            return True
+        except ClientError as e:
+            if e.response["Error"]["Code"] == "ValidationException":
+                self.log.info("Job name '%s' does not exist.", job_name)
+                return False
+            raise e
+
+    def get_job_arn(self, job_name: str) -> str:
+        return self._get_job_by_name(job_name)["jobArn"]
+
+
 class BedrockRuntimeHook(AwsBaseHook):
     """
     Interact with the Amazon Bedrock Runtime.

airflow/providers/amazon/aws/operators/bedrock.py

@@ -19,10 +19,15 @@
 import json
 from typing import TYPE_CHECKING, Any, Sequence
 
-from airflow.providers.amazon.aws.hooks.bedrock import BedrockRuntimeHook
+from airflow.configuration import conf
+from airflow.exceptions import AirflowException
+from airflow.providers.amazon.aws.hooks.bedrock import BedrockHook, BedrockRuntimeHook
 from airflow.providers.amazon.aws.operators.base_aws import AwsBaseOperator
+from airflow.providers.amazon.aws.triggers.bedrock import BedrockCustomizeModelCompletedTrigger
+from airflow.providers.amazon.aws.utils import validate_execute_complete_event
 from airflow.providers.amazon.aws.utils.mixins import aws_template_fields
 from airflow.utils.helpers import prune_dict
+from airflow.utils.timezone import utcnow
 
 if TYPE_CHECKING:
     from airflow.utils.context import Context
@@ -91,3 +96,155 @@ def execute(self, context: Context) -> dict[str, str | int]:
         self.log.info("Bedrock %s prompt: %s", self.model_id, self.input_data)
         self.log.info("Bedrock model response: %s", response_body)
         return response_body
+
+
+class BedrockCustomizeModelOperator(AwsBaseOperator[BedrockHook]):
+    """
+    Create a fine-tuning job to customize a base model.
+
+    .. seealso::
+        For more information on how to use this operator, take a look at the guide:
+        :ref:`howto/operator:BedrockCustomizeModelOperator`
+
+    :param job_name: A unique name for the fine-tuning job.
+    :param custom_model_name: A name for the custom model being created.
+    :param role_arn: The Amazon Resource Name (ARN) of an IAM role that Amazon Bedrock can assume
+        to perform tasks on your behalf.
+    :param base_model_id: Name of the base model.
+    :param training_data_uri: The S3 URI where the training data is stored.
+    :param output_data_uri: The S3 URI where the output data is stored.
+    :param hyperparameters: Parameters related to tuning the model.
+    :param check_if_job_exists: If set to true, operator will check whether a model customization
+        job already exists for the name in the config. (Default: True)
+    :param action_if_job_exists: Behavior if the job name already exists. Options are "timestamp" (default),
+        and "fail"
+    :param customization_job_kwargs: Any optional parameters to pass to the API.
+
+    :param wait_for_completion: Whether to wait for cluster to stop. (default: True)
+    :param waiter_delay: Time in seconds to wait between status checks.
+    :param waiter_max_attempts: Maximum number of attempts to check for job completion.
+    :param deferrable: If True, the operator will wait asynchronously for the cluster to stop.
+        This implies waiting for completion. This mode requires aiobotocore module to be installed.
+        (default: False)
+    :param aws_conn_id: The Airflow connection used for AWS credentials.
+        If this is ``None`` or empty then the default boto3 behaviour is used. If
+        running Airflow in a distributed manner and aws_conn_id is None or
+        empty, then default boto3 configuration would be used (and must be
+        maintained on each worker node).
+    :param region_name: AWS region_name. If not specified then the default boto3 behaviour is used.
+    :param verify: Whether or not to verify SSL certificates. See:
+        https://boto3.amazonaws.com/v1/documentation/api/latest/reference/core/session.html
+    :param botocore_config: Configuration dictionary (key-values) for botocore client. See:
+        https://botocore.amazonaws.com/v1/documentation/api/latest/reference/config.html
+    """
+
+    aws_hook_class = BedrockHook
+    template_fields: Sequence[str] = aws_template_fields(
+        "job_name",
+        "custom_model_name",
+        "role_arn",
+        "base_model_id",
+        "hyperparameters",
+        "check_if_job_exists",
+        "action_if_job_exists",
+        "customization_job_kwargs",
+    )
+
+    def __init__(
+        self,
+        job_name: str,
+        custom_model_name: str,
+        role_arn: str,
+        base_model_id: str,
+        training_data_uri: str,
+        output_data_uri: str,
+        hyperparameters: dict[str, str],
+        check_if_job_exists: bool = True,
+        action_if_job_exists: str = "timestamp",
+        customization_job_kwargs: dict[str, Any] | None = None,
+        wait_for_completion: bool = True,
+        waiter_delay: int = 120,
+        waiter_max_attempts: int = 75,
+        deferrable: bool = conf.getboolean("operators", "default_deferrable", fallback=False),
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.wait_for_completion = wait_for_completion
+        self.waiter_delay = waiter_delay
+        self.waiter_max_attempts = waiter_max_attempts
+        self.deferrable = deferrable
+
+        self.job_name = job_name
+        self.custom_model_name = custom_model_name
+        self.role_arn = role_arn
+        self.base_model_id = base_model_id
+        self.training_data_config = {"s3Uri": training_data_uri}
+        self.output_data_config = {"s3Uri": output_data_uri}
+        self.hyperparameters = hyperparameters
+        self.check_if_job_exists = check_if_job_exists
+        self.customization_job_kwargs = customization_job_kwargs or {}
+        self.action_if_job_exists = action_if_job_exists.lower()
+
+        self.valid_action_if_job_exists: set[str] = {"timestamp", "fail"}
+
+    def execute_complete(self, context: Context, event: dict[str, Any] | None = None) -> str:
+        event = validate_execute_complete_event(event)
+
+        if event["status"] != "success":
+            raise AirflowException(f"Error while running job: {event}")
+
+        self.log.info("Bedrock model customization job `%s` complete.", self.job_name)
+        return self.hook.get_job_arn(event["job_name"])
+
+    def _validate_action_if_job_exists(self):
+        if self.action_if_job_exists not in self.valid_action_if_job_exists:
+            raise AirflowException(
+                f"Invalid value for argument action_if_job_exists {self.action_if_job_exists}; "
+                f"must be one of: {self.valid_action_if_job_exists}."
+            )
+
+    def execute(self, context: Context) -> dict:
+        self._validate_action_if_job_exists()
+
+        if self.check_if_job_exists and self.hook.job_name_exists(self.job_name):
+            if self.action_if_job_exists == "fail":
+                raise AirflowException(f"A Bedrock job with name {self.job_name} already exists.")
+            self.job_name = f"{self.job_name}-{int(utcnow().timestamp())}"
+            self.log.info("Changed job name to '%s' to avoid collision.", self.job_name)
+
+        self.log.info("Creating Bedrock model customization job '%s'.", self.job_name)
+
+        response = self.hook.conn.create_model_customization_job(
+            jobName=self.job_name,
+            customModelName=self.custom_model_name,
+            roleArn=self.role_arn,
+            baseModelIdentifier=self.base_model_id,
+            trainingDataConfig=self.training_data_config,
+            outputDataConfig=self.output_data_config,
+            hyperParameters=self.hyperparameters,
+            **self.customization_job_kwargs,
+        )
+
+        if response["ResponseMetadata"]["HTTPStatusCode"] != 201:
+            raise AirflowException(f"Bedrock model customization job creation failed: {response}")
+
+        task_description = f"Bedrock model customization job {self.job_name} to complete."
+        if self.deferrable:
+            self.log.info("Deferring for %s", task_description)
+            self.defer(
+                trigger=BedrockCustomizeModelCompletedTrigger(
+                    job_name=self.job_name,
+                    waiter_delay=self.waiter_delay,
+                    waiter_max_attempts=self.waiter_max_attempts,
+                    aws_conn_id=self.aws_conn_id,
+                ),
+                method_name="execute_complete",
+            )
+        elif self.wait_for_completion:
+            self.log.info("Waiting for %s", task_description)
+            self.hook.get_waiter("model_customization_job_complete").wait(
+                jobIdentifier=self.job_name,
+                WaiterConfig={"Delay": self.waiter_delay, "MaxAttempts": self.waiter_max_attempts},
+            )
+
+        return response["jobArn"]

airflow/providers/amazon/aws/sensors/bedrock.py

@@ -0,0 +1,111 @@
+#
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you 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.
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Any, Sequence
+
+from airflow.configuration import conf
+from airflow.providers.amazon.aws.sensors.base_aws import AwsBaseSensor
+from airflow.providers.amazon.aws.triggers.bedrock import BedrockCustomizeModelCompletedTrigger
+from airflow.providers.amazon.aws.utils.mixins import aws_template_fields
+
+if TYPE_CHECKING:
+    from airflow.utils.context import Context
+
+from airflow.exceptions import AirflowException, AirflowSkipException
+from airflow.providers.amazon.aws.hooks.bedrock import BedrockHook
+
+
+class BedrockCustomizeModelCompletedSensor(AwsBaseSensor[BedrockHook]):
+    """
+    Poll the state of the model customization job until it reaches a terminal state; fails if the job fails.
+
+    .. seealso::
+        For more information on how to use this sensor, take a look at the guide:
+        :ref:`howto/sensor:BedrockCustomizeModelCompletedSensor`
+
+
+    :param job_name: The name of the Bedrock model customization job.
+
+    :param deferrable: If True, the sensor will operate in deferrable more. This mode requires aiobotocore
+        module to be installed.
+        (default: False, but can be overridden in config file by setting default_deferrable to True)
+    :param max_retries: Number of times before returning the current state, defaults to None
+    :param poke_interval: Polling period in seconds to check for the status of the job.
+    :param aws_conn_id: The Airflow connection used for AWS credentials.
+        If this is ``None`` or empty then the default boto3 behaviour is used. If
+        running Airflow in a distributed manner and aws_conn_id is None or
+        empty, then default boto3 configuration would be used (and must be
+        maintained on each worker node).
+    :param region_name: AWS region_name. If not specified then the default boto3 behaviour is used.
+    :param verify: Whether or not to verify SSL certificates. See:
+        https://boto3.amazonaws.com/v1/documentation/api/latest/reference/core/session.html
+    :param botocore_config: Configuration dictionary (key-values) for botocore client. See:
+        https://botocore.amazonaws.com/v1/documentation/api/latest/reference/config.html
+    """
+
+    INTERMEDIATE_STATES = ("InProgress",)
+    FAILURE_STATES = ("Failed", "Stopping", "Stopped")
+    SUCCESS_STATES = ("Completed",)
+    FAILURE_MESSAGE = "Bedrock model customization job sensor failed."
+
+    aws_hook_class = BedrockHook
+    template_fields: Sequence[str] = aws_template_fields("job_name")
+    ui_color = "#66c3ff"
+
+    def __init__(
+        self,
+        *,
+        job_name: str,
+        max_retries: int = 75,
+        poke_interval: int = 120,
+        deferrable: bool = conf.getboolean("operators", "default_deferrable", fallback=False),
+        **kwargs: Any,
+    ) -> None:
+        super().__init__(**kwargs)
+        self.job_name = job_name
+        self.poke_interval = poke_interval
+        self.max_retries = max_retries
+        self.deferrable = deferrable
+
+    def execute(self, context: Context) -> Any:
+        if self.deferrable:
+            self.defer(
+                trigger=BedrockCustomizeModelCompletedTrigger(
+                    job_name=self.job_name,
+                    waiter_delay=int(self.poke_interval),
+                    waiter_max_attempts=self.max_retries,
+                    aws_conn_id=self.aws_conn_id,
+                ),
+                method_name="poke",
+            )
+        else:
+            super().execute(context=context)
+
+    def poke(self, context: Context) -> bool:
+        state = self.hook.get_customize_model_job_state(self.job_name)
+
+        if state in self.FAILURE_STATES:
+            # TODO: remove this if block when min_airflow_version is set to higher than 2.7.1
+            if self.soft_fail:
+                raise AirflowSkipException(self.FAILURE_MESSAGE)
+            raise AirflowException(self.FAILURE_MESSAGE)
+
+        if state in self.INTERMEDIATE_STATES:
+            return False
+        return True