Re: Very high latency, low bandwidth replication

Hi Bob.

On Mon, Jun 30, 2014 at 10:05 AM, Bob Jolliffe <bobjolliffe@gmail.com> wrote:
> What are people's thoughts about a more optimal solution?  I would like to
> use a more incremental approach to replication.  This does not have to be a
> "live" replication .. asynchronously triggering once every 24 hours is
> sufficient.  Also there are only a subset of tables which are required (the
> rest consist of data which is generated).

If you only need to replicate once every 24 hours, which means you can
tolerate lags, you could try log shipping. Instead of sending the wal
records from master to standby directly just spool them, compress them
as much as you can ( I would try pglesslog plus an XZ on it's output
), and send it once a day. This for the 'incremental part'. For the
only a subset of tables, you could try to set up a local partial
mirror using any of the trigger based replication products and then do
log-shipping of that.

Also, the logical replication slot stuff added to the latest version
seems really promissing for this kind of thing, but I'm not familiar
enough with it to recommend anything.

Also, depending on your data updating patterns, database sizes and
other stuff, a trigger based replication approach can save a lot of
traffic. I mean, if you have records which are heavily updated, but
only replicate once a day, you can collapse all the day stuff in a
single update. I once did a similar thing to transmit deltas over a
2400bps modem by making daily sorted dumps and sending daily deltas
with previous day ( it needed a bit of coding, about a couple hundred
lines, but produced ridiculously small deltas, and with a bit of care
their application was idempotent, which simplified the recovery on
errors ).

   Francisco Olarte.