Re: BUG #16125: Crash of PostgreSQL's wal sender during logical replication

On 2019-Nov-19, Andres Freund wrote:

> > > 2525:rmgr: Heap len (rec/tot): 1671/ 1671, tx: 1667601530, lsn: 25EE/D66F5508, prev 25EE/D66F54C8, desc: INSERT off 2, blkref #0: rel 1663/13018/88964792 blk 18475
> > > 2532:rmgr: Heap len (rec/tot): 54/ 54, tx: 1667601530, lsn: 25EE/D66F7EF0, prev 25EE/D66F7EA0, desc: DELETE off 2 , blkref #0: rel 1663/13018/88964792 blk 18475

> > So, this DELETE change triggers the issue. If I understand it correctly,
> > that means the transaction locked the tuple and then deleted it. But I
> > don't see any record about the lock.
>
> Hm. I don't think it necessarily means that. compute_new_xmax_infomask(), when
> called from heap_delete(), will e.g. set HEAP_KEYS_UPDATED when
>
> if (old_infomask & HEAP_XMAX_INVALID)
>
> is true. Which is the most common case. I actually find it *more*
> curious to *not* see that for a delete. I've not traced fully through
> it, yet.
>
>
> Alvaro, it's not clear to me whether the the multixact codepaths in
> compute_new_xmax_infomask() actually work as expected for deletes.

So, what's really happening here is heap_insert(speculative) followed by
heap_abort_speculative(). That seems to explain the above WAL sequence.
I *think* the speculative insertion code in reorderbuffer simply delays
the tuple insertion until it is confirmed; so this sequence would result
in nothing at all happening at ReorderBufferCommit. Unless something
has messed up the state somehow.

--
Álvaro Herrera https://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services