Customize Logging#

Server Logging#

BentoML provides a powerful and detailed logging pattern out of the box. Request logs for webservices are logged along with requests to each of the model runner services.

The request log format is as follows:

time [LEVEL] [component] ClientIP:ClientPort (scheme,method,path,type,length) (status,type,length) Latency (trace,span,sampled)

For example, a log message might look like:

2022-06-28T18:07:35-0700 [INFO] [api_server] (scheme=http,method=POST,path=/classify,type=application/json,length=20) (status=200,type=application/json,length=3) 0.005ms (trace=67131233608323295915755120473254509377,span=4151694932783368069,sampled=0)

OpenTelemetry Compatible#

The BentoML logging system implements the OpenTelemetry standard for http throughout the call stack to provide for maximum debuggability. Propogation of the OpenTelemetry parameters follows the standard provided here

The following are parameters which are provided in the logs as well for correlation back to particular requests.

  • trace is the id of a trace which tracks “the progression of a single request, as it is handled

    by services that make up an application” - OpenTelemetry Basic Documentation

  • span is the id of a span which is contained within a trace. “A span is the building block of a

    trace and is a named, timed operation that represents a piece of the workflow in the distributed system. Multiple spans are pieced together to create a trace.” - OpenTelemetry Span Documentation

  • sampled is the number of times this trace has been sampled. “Sampling is a mechanism to control

    the noise and overhead introduced by OpenTelemetry by reducing the number of samples of traces collected and sent to the backend.” - OpenTelemetry SDK Documentation

Logging Configuration#

Access logs can be configured by setting the appropriate flags in the bento configuration file for both web requests and model serving requests. Read more about how to use a bento configuration file here in the - Configuration Guide

To configure other logs, use the default python logging configuration. All BentoML logs are logged under the "bentoml" namespace.

Web Service Request Logging#

For web requests, logging can be enabled and disabled using the logging.access parameter at the top level of the bentoml_configuration.yml.

      enabled: False
      # whether to log the size of the request body
      request_content_length: True
      # whether to log the content type of the request
      request_content_type: True
      # whether to log the content length of the response
      response_content_length: True
      # whether to log the content type of the response
      response_content_type: True

Model Runner Request Logging#

Depending on how you’ve configured BentoML, the webserver may be separated from the model runner. In either case, we have special logging that is enabled specifically on the model side of the request. You may configure the runner access logs under the runners parameter at the top level of your bentoml_configuration.yml:

          enabled: True

The available configuration options are identical to the webserver request logging options above. These logs are disabled by default in order to prevent double logging of requests.

Library Logging#

When using BentoML as a library, BentoML does not configure any logs. By default, Python will configure a root logger that logs at level WARNING and higher. If you want to see BentoML’s DEBUG or INFO logs, register a log handler to the bentoml namespace:

import logging

ch = logging.StreamHandler()
formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s')

bentoml_logger = logging.getLogger("bentoml")