This sounds like a cool idea, but mtail doesn't support it at the moment.
The two main issues, and several sub issues I can think of are:
* How does one expect to run mtail in this mode
* is it a one-off command run by the operator on specific logs, or should mtail always reprocess logs that match on startup
* will Prometheus know to discard duplicate data points if you reprocess a file that was partially read already, or does mtail need to manage that state
* how does mtail know if a log file has been read already, and at which byte offset has it started at
* How should mtail sample the data points in order to construct a timeseries, because right now oneshot doesn't report any counters until the end -- one assumes that mtail must export every 10s or whatever to make this work.
Would you mind filing a feature request at http://github.com/google/mtail/issues
with a summary of how you would like to use mtail to backfill data? If you can speculate on command lines and expected behaviour that would be really helpful as well.