[syslog-ng] [RFC]: afsql improvement plans

[Martin Holste]

> I'll read up on it, and probably go with supporting both LOAD DATA and
> HandlerSocket (after thinking a bit more about the INSERT variations,
> the few things I didn't like about LOAD DATA aren't that severe that
> would make it worth exploring the INSERT options).
>
Good points.  LOAD DATA really is the holy grail of INSERT for both
speed and universality.  As I wrote earlier, just writing to /dev/shm
means you don't need tmpfs on Linux.  You might want to look at the
source code for the mysqlimport binary to see how they implement the
various low-level functions for LOAD DATA.  Surely there's a way to
bypass the ridiculousness of a buffer file if you can load straight
from in-program memory in syslog-ng.  When looking at an
implementation, don't forget the functionality that comes with
"CONCURRENT" so that it doesn't need to lock the table or block.  The
manual says there is a performance hit for doing this, but I can
assure you it is unnoticeable.