The great connection cleanup

[zalando]

The great connection cleanup

[alexeyklyukin]

I think we should accept both name=value and URI format when getting connection strings from the user.

As for the internal format, it makes sense to store all connection informations as dicts. This simplifies checking whether two conninfos represent the same connection, as well as modifying them.

As for the primary_conninfo in recovery.conf, we don't have a choice: the URI format was only introduced in 9.5, which is yet to be released. It makes sense to use the name=value connection strings when we output the value somewhere (i.e. to the recovery.conf, pg_rewind connection string, etc).

As for the usage of two types of connections, it won't work: pg_rewind requires a superuser connection to the master host, so you need it as well (there is a discussion to patch pg_rewind to use only replication connection, but this is not going to happen in 9.5).

The idea to go through the API for everything except the local, the replication and pg_rewind connection looks reasonable to me (aren't we doing it already?)

Note that we might be slow on providing feedback this week: the team is on PGConf.EU conference until Monday...

[jberkus]

Ah, ok, so there's three types of connections then. I haven't really looked at the pg_rewind code; I hadn't planned on using pg_rewind short of 9.5.

I'd forgotten about that limitation with recovery.conf. So, then we should make the canonical type dictionaries? With a single canonical function for creating connection strings?

If we're not going to use URIs, then wouldn't it make more sense to store the connection info as JSON in the DCS?

[a1exsh]

On Thu, Oct 29, 2015 at 6:58 PM, Josh Berkus [email protected]
wrote:

Ah, ok, so there's three types of connections then. I haven't really
looked at the pg_rewind code; I hadn't planned on using pg_rewind short of
9.5.

I'd forgotten about that limitation with recovery.conf. So, then we should
make the canonical type dictionaries? With a single canonical function for
creating connection strings?

Actually I wonder why this wasn't back-patched to 9.2... Maybe we should
raise a point on the hackers list. Original discussion:
http://www.postgresql.org/message-id/[email protected]

Alex

[jberkus]

Are we sure it isn't supported? I'll test.

[a1exsh]

On Thu, Oct 29, 2015 at 9:37 PM, Josh Berkus [email protected]
wrote:

Are we sure it isn't supported? I'll test.

I am pretty sure it isn't, because of this:
https://github.com/postgres/postgres/blob/REL9_4_STABLE/src/backend/replication/libpqwalreceiver/libpqwalreceiver.c#L90

Appending that to an URI doesn't do the trick.

[feikesteenbergen]

three types of connections then

The third can be generated from the key value pairs, as you can specify the connection string that way:

$ pg_rewind -D /tmp/ --source-server="dbname=postgres host=localhost user=dummy"

could not connect to server: FATAL:  role "dummy" does not exist

[LappleApple]

@jberkus @feikesteenbergen @a1exsh @alexeyklyukin @CyberDem0n Any agreement/compromise we can come to here so we can adopt/formalize any necessary standards + close out the issue?

[jberkus]

Well, because we didn't do this 4 months ago, it's going to be much harder now ...