Re: Long text values destroys logical replication slots

On 2015-10-28 23:00:53 +0100, Michael Paquier wrote:
> On Wed, Oct 28, 2015 at 10:17 PM, Andres Freund <andres(at)anarazel(dot)de> wrote:
> > Can you reproduce it with test_decoding as the output plugin?
>
> You can just use that for example to get an assertion failure:
> CREATE TABLE a (b text);
> ALTER TABLE ONLY a REPLICA IDENTITY FULL;
> SELECT * FROM pg_create_logical_replication_slot('new', 'test_decoding');
> INSERT INTO a (b) VALUES (repeat('k', 2000000));
> UPDATE a SET b = 'c';
> select * from pg_logical_slot_peek_changes('new', NULL, NULL); -- boom
>
> frame #3: 0x0000000100458ca9
> postgres`DecodeXLogTuple(data=0x00007fb7e2126046, len=22910,
> tuple=0x000000010a32e038) + 137 at decode.c:856
> 853 int datalen = len - SizeOfHeapHeader;
> 854
> 855 Assert(datalen >= 0);
> -> 856 Assert(datalen <= MaxHeapTupleSize);
> (lldb) p datalen
> (int) $0 = 22905

Ugh, that's some monumental stupidity on my side in the handling of
oldtuple values. Newtuple datums never can be bigger than 8k (all
relevant columns point into toasted datums)- but that's absolutely not
the case for the "old" values. That can't happen for primary keys (due
to the btree limitations), which is why this wasn't noticed so far.
Gotta think about this one.

Greetings,

Andres Freund