On 2021-Aug-17, Bossart, Nathan wrote: > On 8/16/21, 5:09 PM, "alvhe...@alvh.no-ip.org" <alvhe...@alvh.no-ip.org> > wrote: > > The reason for the latter is that I suspect the segment boundary map > > will also have a use to fix the streaming replication issue I mentioned > > earlier in the thread. This also makes me think that we'll want the wal > > record *start* address to be in the hash table too, not just its *end* > > address. So we'll use the start-1 as position to send, rather than the > > end-of-segment when GetFlushRecPtr() returns that. > > I've been thinking about the next steps for this one, too. ISTM we'll > need to basically assume that the flush pointer jumps along record > boundaries except for the cross-segment records. I don't know if that > is the safest assumption, but I think the alternative involves > recording every record boundary in the map.
I'm not sure I understand your idea correctly. Perhaps another solution is to assume that the flush pointer jumps along record boundaries *including* for cross-segment records. The problem stems precisely from the fact that we set the flush pointer at segment boundaries, even when they aren't record boundary. -- Álvaro Herrera 39°49'30"S 73°17'W — https://www.EnterpriseDB.com/ Bob [Floyd] used to say that he was planning to get a Ph.D. by the "green stamp method," namely by saving envelopes addressed to him as 'Dr. Floyd'. After collecting 500 such letters, he mused, a university somewhere in Arizona would probably grant him a degree. (Don Knuth)
