Serving: Verbose logging for prediction

[waichee]

Hi

Just wondering is it possible to turn on more verbose logging such it will log out information on the requests being processed?

[kirilg]

You should be able to enable VLOG log statements that are not printed by default (search code for VLOG(1) for example) by setting the environment variable TF_CPP_MIN_VLOG_LEVEL=1. If the existing logging is insufficient, you can also add your own VLOG statements to help in your debugging.

[waichee]

thanks @kirilg any chance that there's also an environment variable to specify the log output format (example to have it spit out json logs)?

[kirilg]

I'm not aware of a way to log the output in any other format, sorry. We use TensorFlow's logging libraries so perhaps you can file a feature request in the TensorFlow repo and they can triage. If they add support, we'll be able to pull that in here as well.

[joekiqin]

it can just print request and response? TF_CPP_MIN_VLOG_LEVEL=1、2、3，print too much log and continuous output