Where I used to work, we had a great system for logging/tracing only sessions we were interested in, while not generating any output for all the other sessions. I wrote about it here: http://henrikwarne.com/2014/01/21/session-based-logging/
One problem is that we run across multiple services, on multiple machines, so we would have to have some way of passing a customer ID into our AZ level logstash forwarders, and find a way of collating ERROR level logs, and sending those transactions up.
We also have the problem of customers asking for retrospective issue debug, so having all the logs is useful.
Am I the only one that stash GBs of low-level log messages in order to diagnose bugs that occurred in the past?