Martijn van Oosterhout <kleptog@svana.org> writes:
> With one big exception: sometimes I/O is non-interruptable (the good
> old D state). In this case the interrupts will happen but will simply
> be queued and in fact will all be dropped except the last one
> (non-realtime signals are never stacked). The handler will probably be
> called the moment it returns to user-space.
If backends store their current status in shared memory then a separate
process entirely can receive the interrupts, scan through the shared memory
process states and do the accounting. It would never be interuppting i/o since
that process would never be doing i/o. It could use real-time signals, reads
on /dev/rtc, or whatever other methods exist for scheduling periodic events
since it would be doing nothing but handling these interrupts.
The neat thing about this is that it would be possible to look at a process
from outside instead of having to hijack the application to get feedback. You
could find information like total time spent in i/o versus cpu aggregated
across all queries from a single backend or for all backends.
The downside is that to handle things like EXPLAIN ANALYZE you would have to
provide enough information about plans to this accounting process for it to
store the information. Plans are currently purely local to the backend running
them.
Perhaps it would be enough to provide a unique id for the plan (sequential id
per backend would be enough). And a unique id for the plan node. The
accounting process wouldn't know anything more about what that node
represented but the backend could later query to fetch the information and
associate it with the plan.
--
greg
