Re: SIGSEGV from START_REPLICATION 0/XXXXXXX in XLogSendPhysical () at walsender.c:2762

On 2020/06/02 13:24, Michael Paquier wrote:
> On Fri, May 29, 2020 at 06:09:06PM +0900, Masahiko Sawada wrote:
>> Yes. Conversely, if we start logical replication in a physical
>> replication connection (i.g. replication=true) we got an error before
>> staring replication:
>>
>> ERROR:  logical decoding requires a database connection
>>
>> I think we can prevent that SEGV in a similar way.
> 
> Still unconvinced as this restriction stands for logical decoding
> requiring a database connection but it is not necessarily true now as
> physical replication has less restrictions than a logical one.

Could you tell me what the benefit for supporting physical replication on
logical rep connection is? If it's only for "undocumented"
backward-compatibility, IMO it's better to reject such "tricky" set up.
But if there are some use cases for that, I'm ok to support that.

> Looking at the code, I think that there is some confusion with the
> fake WAL reader used as base reference in InitWalSender() where we
> assume that it could only be used in the context of a non-database WAL
> sender.  However, this initialization happens when the WAL sender
> connection is initialized, and what I think this misses is that we
> should try to initialize a WAL reader when actually going through a
> START_REPLICATION command.
> 
> I can note as well that StartLogicalReplication() moves in this sense
> by setting xlogreader to be the one from logical_decoding_ctx once the
> decoding context has been created.
> 
> This results in the attached.  The extra test from upthread to check
> that logical decoding is not allowed in a non-database WAL sender is a
> good idea, so I have kept it.

Yes. Also we should add the test to check if physical replication can work
fine even on logical rep connection?

Regards,

-- 
Fujii Masao
Advanced Computing Technology Center
Research and Development Headquarters
NTT DATA CORPORATION