Обсуждение: pg_receivexlog sometimes fails on first start
When starting pg_receivexlog for the first time, i.e. with an empty directory, it sometimes fails like this: pg_receivexlog: unexpected termination of replication stream: ERROR: requested starting point 56/70000000 is ahead of the WAL flush position of this server 56/6FFC4000 This is easier to reproduce if you have a steady stream of updates, with no commits, running concurrently. A bulk load, for example. The problem is that pg_receivexlog issues the IDENTIFY_SYSTEM command, which returns the current WAL insert position. It then requests to start streaming from that position. Or to be precise, from the beginning of the WAL segment containing that position. That's wrong; the WAL might not be flushed up to the insert position yet. I think returning the insert position in IDENTIFY_SYSTEM is a bad idea. We even mention in the documentation that that field is "useful to get a known location in the transaction log where streaming can start", but the insert position is wrong for that purpose. Any objections to changing IDENTIFY_SYSTEM to return the flush position, instead of insert position? We haven't heard any complaints from the field about this, but IMHO that should also be back-patched. It's a design bug, and the fix is simple and unlikely to cause any harm. - Heikki
On Fri, Feb 6, 2015 at 6:17 AM, Heikki Linnakangas <hlinnakangas@vmware.com> wrote: > When starting pg_receivexlog for the first time, i.e. with an empty > directory, it sometimes fails like this: > > pg_receivexlog: unexpected termination of replication stream: ERROR: > requested starting point 56/70000000 is ahead of the WAL flush position of > this server 56/6FFC4000 > > This is easier to reproduce if you have a steady stream of updates, with no > commits, running concurrently. A bulk load, for example. > > The problem is that pg_receivexlog issues the IDENTIFY_SYSTEM command, which > returns the current WAL insert position. It then requests to start streaming > from that position. Or to be precise, from the beginning of the WAL segment > containing that position. That's wrong; the WAL might not be flushed up to > the insert position yet. > > I think returning the insert position in IDENTIFY_SYSTEM is a bad idea. We > even mention in the documentation that that field is "useful to get a known > location in the transaction log where streaming can start", but the insert > position is wrong for that purpose. Good catch. > Any objections to changing IDENTIFY_SYSTEM to return the flush position, > instead of insert position? We haven't heard any complaints from the field > about this, but IMHO that should also be back-patched. It's a design bug, > and the fix is simple and unlikely to cause any harm. +1 with your idea. Regards, -- Fujii Masao
On Fri, Feb 6, 2015 at 4:02 AM, Fujii Masao wrote: > On Fri, Feb 6, 2015 at 6:17 AM, Heikki Linnakangas wrote >> Any objections to changing IDENTIFY_SYSTEM to return the flush position, >> instead of insert position? We haven't heard any complaints from the field >> about this, but IMHO that should also be back-patched. It's a design bug, >> and the fix is simple and unlikely to cause any harm. > > +1 with your idea. +1. -- Michael