Hi
Since my attempts to find a simple solution for the read-only query
locking problems (Once that doesn't need full wal logging of lock
requests) haven't been successfully yet, I've decided to turn to the
problems of tracking a snapshot on the slaves for now. (Because first
such a snapshot is needed for any kind of concurrent recovery anyway, and
second because any non-simplistic solution of the locking problems
will quite likely benefit from such a snapshot).
The idea is to create a special kind of snapshot that works basically
like a MVCC snapshot, but with the meaning of the xip array inverted.
Usually, if a xid is *not* in the xip array of a snapshot, and greater
than the xmin of that snapshot, the clog state of the xid determines
tuple visibility. This is not well suited for queries running during
replay, because the effects of a xlog record with a (to the slave)
previously unknown xid shouldn't be visible to concurrently running
queries.
Therefore, flag xip_inverted will be added to SnapshotData, that causes
HeapTupleSatisfiesMVCC to assume that any xid >= xmin and *not* in the
xip array is in progress.
This allows the following to work:
.) Store RecentXmin with every xlog record, in a new field xl_xmin. (Wouldn't be needed in *every* record, but for
nowkeeping it directly inside XLogRecord make things easier, and it's just 4 bytes)
.) Maintain a global snapshot template in shmem during replay, with the xmin being the highest xmin seen so far in
anyxlog record. That template is copied whenever a readonly query needs to obtain a snapshot.
.) Upon replaying a COMMIT or COMMIT_PREPARED record, the xmin of the to-be-committed transaction is added to the
globalsnapshot, making the commit visibile to all further copies of that snapshot.
If you can shoot this down, you're welcome to do so ;-)
greetings, Florian Pflug
