About a week ago, I wrote about moving a piece of software out of writing
multiple files and into LMDB.
Currently, the software will run on UBIFS (which is unsafe for NOSYNC
option). Is multi-threaded and has no concept of transactions. Write
operations are spread between threads. It only uses a mutex to avoid
multiple writers, and tend to happen in a burst. Storage is implemented by
a layer that has a similar API (read, write, delete) to LMDB.
My main goal is to replace the guts of this layer with LMDB (while keeping
the API), to get an idea on how the program would work. Because of the lack
of transaction concept in the current API, my initial approach was to
accumulate transactions (after the first sync) for a couple of seconds, and
then call env_sync. This solution is not power failure robust because of
I was trying to figure out other solutions that can be put this deferred
concept in place, again, with the goal of proving worthiness of a full
refactoring of the code in order to properly use LMDB.
Show replies by date