Martijn van Oosterhout <kleptog@svana.org> writes:
> On Fri, Oct 13, 2006 at 03:35:37PM -0000, Andrew - Supernews wrote:
>> It's just the number of disk revolutions per second. Without caching, each
>> WAL flush tends to require a whole revolution unless the on-disk layout of
>> the filesystem is _very_ strange.
> Is that really true? In theory block n+1 could be half a revolution
> after block n, allowing you to commit two transactions per revolution.
Not relevant, unless the prior transaction happened to end exactly at a
WAL block boundary. Otherwise, you still have to re-write the back end
of the same disk block the previous transaction wrote into. (In
practice, for the sort of tiny transactions that are at stake here,
quite a few xacts fit into a single WAL block so the same block is
rewritten several times before moving on to the next.)
There was a long thread in -hackers a couple years back exploring ways
to break this "1 xact per disk rotation" barrier with more creative
layouts of the WAL files, but nobody could come up with something that
looked reasonably robust --- ie, both safe and not full of unsupportable
assumptions about knowing exactly where everything actually is on the
disk platter. It'd still be interesting if anyone gets a new idea...
regards, tom lane