Server is unable to recover when disk log is incomplete due to a crash while writing an entry

[palmskog]

From @pfons on April 13, 2016 0:17

A crash during a write to the disk log (by M.to_channel in function save on file Shim.ml) can cause a partial entry to be appended to the log. When this happens the server that crashed is not able to recover and produces the following error when starting:
"Fatal error: exception Failure("input_value: truncated object")"

This problem can be simulated by stopping the server and subsequently trimming the last byte from the log. We expect partial writes during crashes to be more likely to occur when the log write cross disk block boundaries.

Copied from original issue: uwplse/verdi#38