Configurations

BentoML provides a configuration interface that allows you to customize the runtime behavior for individual Services within a Bento. This document explains the available configuration fields and offers best practice recommendations for configuring your BentoML Services.

How do configurations work

BentoML’s default configurations are suitable for a wide range of use cases. However, for more granular control over BentoML’s features, you can customize these runtime behaviors (like resource allocation and timeout) using the @bentoml.service decorator for each Service in your service.py file.

Note

If you are using BentoML versions prior to 1.2, you need to set these runtime configurations via a separate configuration.yaml file.

You only need to specify the configurations you want to customize. BentoML automatically fills in any unspecified configurations with their default values. The following is an example:

@bentoml.service(
    resources={"cpu": "2"},
    traffic={"timeout": 60},
)
class Summarization:
    # Service implementation
  • "cpu": "2" indicates that the Service is allocated 2 CPU cores. This is important for ensuring the Service has sufficient processing resources.

  • "timeout": 60 sets the maximum duration (in seconds) that the Service will wait for a response before timing out. In this case, it’s set to 60 seconds.

Configuration fields

BentoML offers a comprehensive set of configuration fields, allowing detailed customization of Services. You can use them to meet the specific requirements of different deployment environments and use cases.

resources

The resources field in BentoML allows you to specify the resource allocation for a Service, including CPU, memory, and GPU. It is useful for managing and optimizing resource usage, particularly when dealing with resource-intensive tasks. Note that this field only takes effect on BentoCloud. Available fields are:

  • cpu: The number (or the amount) of CPUs that the Service should use. This is a string, like “200m” or “1”.

  • memory: The amount of memory that the Service should use.

  • gpu: The number of GPUs allocated to the Service.

  • gpu_type: A specific type of GPU to be used on BentoCloud, following the naming conventions of cloud providers like AWS EC2 or GCP GPU instances. Available values include nvidia-tesla-t4, nvidia-tesla-a100, nvidia-a100-80gb, nvidia-a10g, nvidia-l4, nvidia-tesla-v100, nvidia-tesla-p100, nvidia-tesla-k80, and nvidia-tesla-p4.

We recommend you specify at least one field of resources, so that resources can be automatically allocated to the Service on BentoCloud.

Here is an example:

@bentoml.service(
    resources={
        "cpu": "1",
        "memory": "500Mi",
        "gpu": 4,
        "gpu_type": "nvidia-tesla-a100"
    }
)
class MyService:
    # Service implementation

workers

Parallelize requests handling are the processes that run the code logic within a Service instance. You use workers in the @bentoml.service decorator to define the process-level parallelism within a Service. This configuration is useful for optimizing performance, particularly for high-throughput or compute-intensive Services. workers defaults to 1.

To specify the number of workers (for example, 3) within a Service:

@bentoml.service(workers=3)
class MyService:
    # Service implementation

workers also allows for dynamic CPU allocation using the cpu_count option. This feature can be particularly useful in environments where you want to automatically scale the number of worker processes based on the available CPU cores.

@bentoml.service(workers="cpu_count")
class MyService:
    # Service implementation

traffic

traffic in BentoML allows you to manage how requests are handled by your Service. It includes settings for managing request concurrency and ensuring timely responses, which are helpful for optimizing the Service’s responsiveness and load management. The following fields are available:

  • timeout: Determines the maximum time the Service will wait for a response to be sent back to the client. The default timeout is set to 60 seconds.

  • max_concurrency: Specifies the hard limit on the number of requests that can be processed simultaneously by a single Service instance. It helps you control the load and prevent the Service from being overwhelmed by too many simultaneous requests.

  • concurrency: A BentoCloud-specific field that represents the ideal number of simultaneous requests a Service is designed to handle. Concurrency helps optimize resource utilization and influences how BentoCloud autoscales your Service. By default, concurrency is set to allow infinite requests to avoid system bottlenecks. For detailed information, see Concurrency and autoscaling.

  • external_queue: A BentoCloud-specific field. When deploying a Service on BentoCloud with this field enabled, an external request queue is used to manage incoming traffic more effectively. This is done by queuing excess requests until they can be processed within the defined concurrency limits.

Here is an example of configuring these settings in your Service definition:

@bentoml.service(
    traffic={
        "timeout": 120,
        "max_concurrency": 50,
        "concurrency": 32, # BentoCloud only
        "external_queue": True, # BentoCloud only
    }
)
class MyService:
    # Service implementation

runner_probe

Configure health check settings on BentoCloud for the Service using the endpoints readyz, livez, and healthz. Available fields are:

  • enabled: Determines whether the health checks are enabled.

  • timeout: The maximum time in seconds to wait for a health check probe to complete before considering it failed.

  • period: The frequency, in seconds, at which the health check probes are performed.

Here is an example:

@bentoml.service(runner_probe={"enabled": True, "timeout": 1, "period": 10})
class MyService:
    # Service implementation

logging

Customize server-side logging, including the content type and length of requests and responses, and trace ID formats.

Here is an example:

@bentoml.service(logging={
    "access": {
        "enabled": True,
        "request_content_length": True,
        "request_content_type": True,
        "response_content_length": True,
        "response_content_type": True,
        "format": {
            "trace_id": "032x",
            "span_id": "016x"
        }
    }
})
class MyService:
    # Service implementation

For more information, see Logging.

ssl

ssl enables SSL/TLS for secure communication over HTTP requests. It is helpful for protecting sensitive data in transit and ensuring secure connections between clients and your Service.

BentoML parses all the available fields directly to Uvicorn. Here is an example:

@bentoml.service(ssl={
    "enabled": True,
    "certfile": "/path/to/certfile",
    "keyfile": "/path/to/keyfile",
    "ca_certs": "/path/to/ca_certs",
    "keyfile_password": "",
    "version": 17,
    "cert_reqs": 0,
    "ciphers": "TLSv1"
})
class MyService:
    # Service implementation

http

http allows you to customize the settings for the HTTP server that serves your BentoML Service.

By default, BentoML starts an HTTP server on port 3000. To change the port:

@bentoml.service(http={"port": 5000})
class MyService:
    # Service implementation

You can configure CORS settings if your Service needs to accept cross-origin requests. By default, CORS is disabled. If it is enabled, all fields under http.cors will be parsed to CORSMiddleware. Here is an example:

@bentoml.service(http={
    "cors": {
        "enabled": True,
        "access_control_allow_origins": ["http://myorg.com:8080", "https://myorg.com:8080"],
        "access_control_allow_methods": ["GET", "OPTIONS", "POST", "HEAD", "PUT"],
        "access_control_allow_credentials": True,
        "access_control_allow_headers": ["*"],
        "access_control_allow_origin_regex": "https://.*\.my_org\.com",
        "access_control_max_age": 1200,
        "access_control_expose_headers": ["Content-Length"]
    }
})
class MyService:
    # Service implementation

Configuring CORS is important when your Service is accessed from web applications hosted on different domains. Proper CORS settings ensure that your Service can securely handle requests from allowed origins, enhancing both security and usability.

By customizing the http configuration, you can fine-tune how your BentoML Service interacts over HTTP, including adapting to specific network environments, securing cross-origin interactions, and ensuring compatibility with various client applications.

monitoring

monitoring allows you to collect logs and keep track of the performance and health of a Service for maintaining its reliability and efficiency. By default, BentoML provides a built-in monitoring mechanism, while you can customize it by setting a configuration file in YAML.

Here is an example:

@bentoml.service(monitoring={
    "enabled": True,
    "type": "default",
    "options": {
        "log_config_file": "path/to/log_config.yaml", # A configuration file for customizing monitoring behavior, using Python's logging module
        "log_path": "monitoring" # The directory where logs will be exported
    }
})
class MyService:
    # Service implementation

For more information, see Monitoring.

metrics

metrics allows you to collect and customize metrics of Counter, Histogram, Summary, and Gauge types. By default, this feature is enabled.

Here is an example:

@bentoml.service(metrics={
    "enabled": True,
    "namespace": "bentoml_service",
    "duration": {
        "buckets": [0.1, 0.2, 0.5, 1, 2, 5, 10]
    }
})
class MyService:
    # Service implementation

For more information, see Metrics.

tracing

You can configure tracing with different exporters like Zipkin, Jaeger, and OTLP. The specific configurations may vary depending on the exporter type defined.

Here is an example:

import bentoml

@bentoml.service(
    resources={"cpu": "2"},
    traffic={"timeout": 10},
    tracing={
        # Common configurations
        "exporter_type": "jaeger",
        "sample_rate": 1.0,
        "timeout": 5,
        "max_tag_value_length": 256,
        "excluded_urls": "readyz",
        "jaeger": {
            # Specific configurations of the exporter
    }
)
class MyService:
   # Service implementation code

For more information, see Tracing.

For full schema of the configurations, see this file.