Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Amazon Bedrock - Model Customization Jobs #38693

Merged
merged 12 commits into from
Apr 8, 2024
Merged
20 changes: 20 additions & 0 deletions airflow/providers/amazon/aws/hooks/bedrock.py
Original file line number Diff line number Diff line change
Expand Up @@ -19,6 +19,26 @@
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)


class BedrockRuntimeHook(AwsBaseHook):
"""
Interact with the Amazon Bedrock Runtime.
Expand Down
161 changes: 160 additions & 1 deletion airflow/providers/amazon/aws/operators/bedrock.py
Original file line number Diff line number Diff line change
Expand Up @@ -19,10 +19,17 @@
import json
from typing import TYPE_CHECKING, Any, Sequence

from airflow.providers.amazon.aws.hooks.bedrock import BedrockRuntimeHook
from botocore.exceptions import ClientError

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
Expand Down Expand Up @@ -91,3 +98,155 @@ 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 ensure_unique_job_name: If set to true, operator will check whether a model customization
job already exists for the name in the config and append the current timestamp if there is a
name conflict. (Default: True)
: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. (default: 120)
:param waiter_max_attempts: Maximum number of attempts to check for job completion. (default: 75)
: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",
"ensure_unique_job_name",
"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],
ensure_unique_job_name: bool = True,
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.ensure_unique_job_name = ensure_unique_job_name
self.customization_job_kwargs = customization_job_kwargs or {}

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.conn.get_model_customization_job(jobIdentifier=event["job_name"])["jobArn"]

def execute(self, context: Context) -> dict:
response = {}
retry = True
while retry:
# If there is a name conflict and ensure_unique_job_name is True, append the current timestamp
# to the name and retry until there is no name conflict.
# - Break the loop when the API call returns success.
# - If the API returns an exception other than a name conflict, raise that exception.
# - If the API returns a name conflict and ensure_unique_job_name is false, raise that exception.
try:
# Ensure the loop is executed at least once, and not repeat unless explicitly set to do so.
retry = False
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,
)
except ClientError as error:
if error.response["Error"]["Message"] != "The provided job name is currently in use.":
raise error
if not self.ensure_unique_job_name:
raise error
retry = True
self.job_name = f"{self.job_name}-{int(utcnow().timestamp())}"
self.log.info("Changed job name to '%s' to avoid collision.", self.job_name)

if response["ResponseMetadata"]["HTTPStatusCode"] != 201:
raise AirflowException(f"Bedrock model customization job creation failed: {response}")
Taragolis marked this conversation as resolved.
Show resolved Hide resolved

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"]
110 changes: 110 additions & 0 deletions airflow/providers/amazon/aws/sensors/bedrock.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,110 @@
#
# 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 mode. 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. (default: 75)
:param poke_interval: Polling period in seconds to check for the status of the job. (default: 120)
: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.conn.get_model_customization_job(jobIdentifier=self.job_name)["status"]
self.log.info("Job '%s' state: %s", self.job_name, state)

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)

return state not in self.INTERMEDIATE_STATES
61 changes: 61 additions & 0 deletions airflow/providers/amazon/aws/triggers/bedrock.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,61 @@
# 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

from airflow.providers.amazon.aws.hooks.bedrock import BedrockHook
from airflow.providers.amazon.aws.triggers.base import AwsBaseWaiterTrigger

if TYPE_CHECKING:
from airflow.providers.amazon.aws.hooks.base_aws import AwsGenericHook


class BedrockCustomizeModelCompletedTrigger(AwsBaseWaiterTrigger):
"""
Trigger when a Bedrock model customization job is complete.

:param job_name: The name of the Bedrock model customization job.
:param waiter_delay: The amount of time in seconds to wait between attempts. (default: 120)
:param waiter_max_attempts: The maximum number of attempts to be made. (default: 75)
:param aws_conn_id: The Airflow connection used for AWS credentials.
"""

def __init__(
self,
*,
job_name: str,
waiter_delay: int = 120,
waiter_max_attempts: int = 75,
aws_conn_id: str | None = None,
) -> None:
super().__init__(
serialized_fields={"job_name": job_name},
ferruzzi marked this conversation as resolved.
Show resolved Hide resolved
waiter_name="model_customization_job_complete",
waiter_args={"jobIdentifier": job_name},
failure_message="Bedrock model customization failed.",
status_message="Status of Bedrock model customization job is",
status_queries=["status"],
return_key="job_name",
return_value=job_name,
waiter_delay=waiter_delay,
waiter_max_attempts=waiter_max_attempts,
aws_conn_id=aws_conn_id,
)

def hook(self) -> AwsGenericHook:
return BedrockHook(aws_conn_id=self.aws_conn_id)
Loading