Re: Stopping logical replication protocol

On 6 May 2016 at 23:23, Vladimir Gordiychuk <folyga(at)gmail(dot)com> wrote:

> I prepare small patch that fix problems describe below. Now *WalSndWriteData
> *first check message from consumer and during decode transaction check
> that replication still active. KeppAlive message now not send if was get
> CopyDone package(keep alive now not necessary we preparing to complete). *WalSndWaitForWal
> *after get CopyDone exit from loop. With apply this patch I get next
> measurements
>

I'll review this, but based on the description and concept I agree it's
useful.

I have been frustrated in the past by the inability to terminate the
logical replication stream and return to command mode without a
disconnect/reconnect so I'd like to have this.

AFAIK there's no particular reason we can't do it, it's just not
implemented. It does have to be a clean exit though, where the logical
decoding context is destroyed, the xact it was running in is closed, etc.
Like for the SQL interface.

You still won't want to do it too often because there's a cost to stopping
decoding and restarting it. We have to re-read from the restart_lsn and
reassemble transactions again up to the point where we can start sending
them to the client again. Especially if you're most of the way through
decoding a big xact, or if there's a long-running xact holding restart_lsn
down this might cost a bit. But no worse than repeatedly calling the SQL
logical decoding interface.

--
Craig Ringer http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services