Re: Patch: Implement failover on libpq connect level.

On Tue, Nov 15, 2016 at 9:42 AM, Alvaro Herrera
<alvherre@2ndquadrant.com> wrote:
> Robert Haas wrote:
>> On Mon, Nov 14, 2016 at 2:38 AM, Mithun Cy <mithun.cy@enterprisedb.com> wrote:
>> > I have not tested with logical replication. Currently we identify the
>> > primary to connect based on result of "SELECT pg_is_in_recovery()". So I
>> > think it works. Do you want me test a particular setup?
>>
>> If logical replication is in use, none of the servers involved would
>> be in recovery.  I'm not sure what command would need to be used to
>> assess whether we've got a master or a standby, but probably not that
>> one.  This gets at one of my earlier complaints about this part of the
>> functionality, which is that hardcoding that particular SQL statement
>> into libpq seems like a giant hack.  However, I'm not sure what to do
>> about it.  The functionality is clearly useful, because JDBC has it,
>> and Victor proposed this patch to add it to libpq, and - totally
>> independently of any of that - EnterpriseDB has a customer who has
>> requested libpq support for this as well.  So I am tempted to just
>> hold my nose and hard-code the SQL as JDBC is presumably already
>> doing.  If we figure out what the equivalent for logical replication
>> would be we can add something to cover that case, too.  It's ugly, but
>> I don't have a better idea, and I think there's value in being
>> compatible with what JDBC has already done (even if it's not what we
>> would have chosen to do tabula rasa).
>
> I would rather come up with something that works in both cases that we
> can extend internally later, say pg_is_primary_node() or something like
> that instead; and we implement it initially by returning the inverse of
> pg_is_in_recovery() for replicated non-logical flocks, while we figure
> out what to do in logical replication.  Otherwise it will be harder to
> change later if we embed it in libpq, and we may be forced into
> supporting nonsensical situations such as having pg_is_in_recovery()
> return true for logical replication primary nodes.

I don't think we'll be backed into a corner like that, because we can
always make this contingent on server version.  libpq will have that
available.

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company