GiST concurrency
Now I basically finished recovery for GiST (of course, it's need a hard testing)
and go to concurrency. As it described in Kornaker, Mohan and Hellerstein's
paper
(http://www.sai.msu.su/~megera/postgres/gist/papers/concurrency/sigmod97-gist.pdf)
it's need a way to get global LSN, in our case - XLogRecPtr of last changed
page. As I understand, I can't use ProcLastRecPtr because it is one-process
wide, I need value stored in shared memory. So, may I add method to xlog.c like
this:
/*
* The returning recptr is the beginning of the current record to fill.
* This value is already stored as LSN for changed data pages.
*/
XLogRecPtr
GetCurrentRecPtr(void) {
XLogCtlInsert *Insert = &XLogCtl->Insert;
XLogRecPtr RecPtr;
LWLockAcquire(WALInsertLock, LW_SHARED);
INSERT_RECPTR(RecPtr, Insert, Insert->curridx);
LWLockRelease(WALInsertLock);
return RecPtr;
}
--
Teodor Sigaev E-mail: teodor@sigaev.ru
WWW: http://www.sigaev.ru/
Teodor Sigaev <teodor@sigaev.ru> writes:
Now I basically finished recovery for GiST (of course, it's need a hard testing)
and go to concurrency. As it described in Kornaker, Mohan and Hellerstein's
paper
(http://www.sai.msu.su/~megera/postgres/gist/papers/concurrency/sigmod97-gist.pdf)
it's need a way to get global LSN, in our case - XLogRecPtr of last changed
page. As I understand, I can't use ProcLastRecPtr because it is one-process
wide, I need value stored in shared memory.
If the method needs a truly global LSN, then it is broken --- the only
way you could have such a value and have it stay good long enough to do
anything with it is to block all other backends from inserting any new
WAL records. Which is the very antithesis of concurrency.
I think you probably misunderstood the paper. It looks to me like the
proposal in the paper is to use the LSN assigned to the WAL record that
represents a page split operation. Which you get from the XLogInsert
--- there's no need for an extra call.regards, tom lane
If the method needs a truly global LSN, then it is broken --- the only
way you could have such a value and have it stay good long enough to do
anything with it is to block all other backends from inserting any new
WAL records. Which is the very antithesis of concurrency.
Global LSN needs to recognize page split produced another process by search
algorithm, no more.
I think you probably misunderstood the paper. It looks to me like the proposal in the paper is to use the LSN assigned to the WAL record that represents a page split operation. Which you get from the XLogInsert --- there's no need for an extra call.
You partially right, I don't read it with care chaper 10.1 last paragraph :(
<quotation>
To alleviate the traffic on this high-frequency counter (LSN - teodor),
descending operations can memorize the node's LSN instead.
</quotation>
So, value of global LSN isn't needed.
--
Teodor Sigaev E-mail: teodor@sigaev.ru
WWW: http://www.sigaev.ru/