vincbeck commented on code in PR #48468:
URL: https://github.com/apache/airflow/pull/48468#discussion_r2023573804
##########
providers/amazon/src/airflow/providers/amazon/aws/operators/bedrock.py:
##########
@@ -869,3 +870,121 @@ def execute(self, context: Context) -> Any:
self.log.info("\nQuery: %s\nRetrieved: %s", self.retrieval_query,
result["retrievalResults"])
return result
+
+
+class BedrockBatchInferenceOperator(AwsBaseOperator[BedrockHook]):
+ """
+ Create a batch inference job to invoke a model on multiple prompts.
+
+ .. seealso::
+ For more information on how to use this operator, take a look at the
guide:
+ :ref:`howto/operator:BedrockBatchInferenceOperator`
+
+ :param job_name: A name to give the batch inference job. (templated)
+ :param role_arn: The ARN of the IAM role with permissions to create the
knowledge base. (templated)
+ :param model_id: Name or ARN of the model to associate with this
provisioned throughput. (templated)
+ :param input_uri: The S3 location of the input data. (templated)
+ :param output_uri: The S3 location of the output data. (templated)
+ :param invoke_kwargs: Additional keyword arguments to pass to the API
call. (templated)
+
+ :param wait_for_completion: Whether to wait for cluster to stop. (default:
True)
+ NOTE: The way batch inference jobs work, your jobs are added to a
queue and done "eventually"
+ so using deferrable mode is much more practical than using
wait_for_completion.
+ :param waiter_delay: Time in seconds to wait between status checks.
(default: 60)
+ :param waiter_max_attempts: Maximum number of attempts to check for job
completion. (default: 10)
+ :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",
+ "role_arn",
+ "model_id",
+ "input_uri",
+ "output_uri",
+ "invoke_kwargs",
+ )
+
+ def __init__(
+ self,
+ job_name: str,
+ role_arn: str,
+ model_id: str,
+ input_uri: str,
+ output_uri: str,
+ invoke_kwargs: dict[str, Any] | None = None,
+ wait_for_completion: bool = True,
+ waiter_delay: int = 60,
+ waiter_max_attempts: int = 10,
+ deferrable: bool = conf.getboolean("operators", "default_deferrable",
fallback=False),
+ **kwargs,
+ ):
+ super().__init__(**kwargs)
+ self.job_name = job_name
+ self.role_arn = role_arn
+ self.model_id = model_id
+ self.input_uri = input_uri
+ self.output_uri = output_uri
+ self.invoke_kwargs = invoke_kwargs or {}
+
+ self.wait_for_completion = wait_for_completion
+ self.waiter_delay = waiter_delay
+ self.waiter_max_attempts = waiter_max_attempts
+ self.deferrable = deferrable
+
+ self.activity = "Bedrock batch inference job"
+
+ def execute_complete(self, context: Context, event: dict[str, Any] | None
= None) -> str:
+ validated_event = validate_execute_complete_event(event)
+
+ if validated_event["status"] != "success":
+ raise AirflowException(f"Error while running {self.activity}:
{validated_event}")
+
+ self.log.info("%s '%s' complete.", self.activity, self.job_name)
+
+ return validated_event["job_arn"]
+
+ def execute(self, context: Context) -> str:
+ response = self.hook.conn.create_model_invocation_job(
+ jobName=self.job_name,
+ roleArn=self.role_arn,
+ modelId=self.model_id,
+ inputDataConfig={"s3InputDataConfig": {"s3Uri": self.input_uri}},
+ outputDataConfig={"s3OutputDataConfig": {"s3Uri":
self.output_uri}},
+ **self.invoke_kwargs,
+ )
+ job_arn = response["jobArn"]
+ self.log.info("%s '%s' started with ARN: %s", self.activity,
self.job_name, job_arn)
+
+ task_description = f"for {self.activity} '{self.job_name}' to
complete."
+ if self.deferrable:
Review Comment:
This should be in `if self.wait_for_completion` no? If a user set
`wait_for_completion` to False, IMO we should not wait regardless of the mode.
What do you think?
##########
providers/amazon/src/airflow/providers/amazon/aws/operators/bedrock.py:
##########
@@ -869,3 +870,121 @@ def execute(self, context: Context) -> Any:
self.log.info("\nQuery: %s\nRetrieved: %s", self.retrieval_query,
result["retrievalResults"])
return result
+
+
+class BedrockBatchInferenceOperator(AwsBaseOperator[BedrockHook]):
+ """
+ Create a batch inference job to invoke a model on multiple prompts.
+
+ .. seealso::
+ For more information on how to use this operator, take a look at the
guide:
+ :ref:`howto/operator:BedrockBatchInferenceOperator`
+
+ :param job_name: A name to give the batch inference job. (templated)
+ :param role_arn: The ARN of the IAM role with permissions to create the
knowledge base. (templated)
+ :param model_id: Name or ARN of the model to associate with this
provisioned throughput. (templated)
+ :param input_uri: The S3 location of the input data. (templated)
+ :param output_uri: The S3 location of the output data. (templated)
+ :param invoke_kwargs: Additional keyword arguments to pass to the API
call. (templated)
+
+ :param wait_for_completion: Whether to wait for cluster to stop. (default:
True)
+ NOTE: The way batch inference jobs work, your jobs are added to a
queue and done "eventually"
+ so using deferrable mode is much more practical than using
wait_for_completion.
+ :param waiter_delay: Time in seconds to wait between status checks.
(default: 60)
+ :param waiter_max_attempts: Maximum number of attempts to check for job
completion. (default: 10)
+ :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",
+ "role_arn",
+ "model_id",
+ "input_uri",
+ "output_uri",
+ "invoke_kwargs",
+ )
+
+ def __init__(
+ self,
+ job_name: str,
+ role_arn: str,
+ model_id: str,
+ input_uri: str,
+ output_uri: str,
+ invoke_kwargs: dict[str, Any] | None = None,
+ wait_for_completion: bool = True,
+ waiter_delay: int = 60,
+ waiter_max_attempts: int = 10,
+ deferrable: bool = conf.getboolean("operators", "default_deferrable",
fallback=False),
+ **kwargs,
+ ):
+ super().__init__(**kwargs)
+ self.job_name = job_name
+ self.role_arn = role_arn
+ self.model_id = model_id
+ self.input_uri = input_uri
+ self.output_uri = output_uri
+ self.invoke_kwargs = invoke_kwargs or {}
+
+ self.wait_for_completion = wait_for_completion
+ self.waiter_delay = waiter_delay
+ self.waiter_max_attempts = waiter_max_attempts
+ self.deferrable = deferrable
+
+ self.activity = "Bedrock batch inference job"
+
+ def execute_complete(self, context: Context, event: dict[str, Any] | None
= None) -> str:
+ validated_event = validate_execute_complete_event(event)
+
+ if validated_event["status"] != "success":
+ raise AirflowException(f"Error while running {self.activity}:
{validated_event}")
+
+ self.log.info("%s '%s' complete.", self.activity, self.job_name)
Review Comment:
Code in `execute_complete` should be stateless. This function might not be
executed where the operator has been instantiated. If `self.job_name` is
something random, then it will generate a new job_name here. In other words,
any piece of information used in `execute_complete` should be extracted from
`event` (unless this is a constant like self.activity)
##########
providers/amazon/src/airflow/providers/amazon/aws/waiters/bedrock.json:
##########
@@ -68,6 +68,140 @@
"state": "failure"
}
]
+ },
+ "batch_inference_complete": {
+ "delay": 120,
+ "maxAttempts": 75,
+ "operation": "GetModelInvocationJob",
+ "acceptors": [
+ {
+ "matcher": "path",
+ "argument": "status",
+ "expected": "Completed",
+ "state": "success"
+ },
+ {
+ "matcher": "path",
+ "argument": "status",
+ "expected": "Failed",
+ "state": "failure"
+ },
+ {
+ "matcher": "path",
+ "argument": "status",
+ "expected": "Stopped",
+ "state": "failure"
+ },
+ {
+ "matcher": "path",
+ "argument": "status",
+ "expected": "PartiallyCompleted",
+ "state": "failure"
+ },
+ {
+ "matcher": "path",
+ "argument": "status",
+ "expected": "Expired",
+ "state": "failure"
+ },
+ {
+ "matcher": "path",
+ "argument": "status",
+ "expected": "Submitted",
+ "state": "retry"
+ },
+ {
+ "matcher": "path",
+ "argument": "status",
+ "expected": "InProgress",
+ "state": "retry"
+ },
+ {
+ "matcher": "path",
+ "argument": "status",
+ "expected": "Stopping",
+ "state": "retry"
Review Comment:
If `stopped` is failure, should we consider this one as failure as well? To
fail fast. This could be an assumption which can lead to some other issues so
feel free to disagree :)
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
