On 22/03/16 07:32, konstantin knizhnik wrote:

On Mar 21, 2016, at 4:30 PM, Petr Jelinek wrote:

On 21/03/16 14:25, Andres Freund wrote:
On 2016-03-21 14:18:27 +0100, Petr Jelinek wrote:
On 21/03/16 14:15, Andres Freund wrote:
Only when the origin is actually setup for the current session. You
to call the replorigin_advance yourself from your apply code.

That's problematic from a durability POV.

Huh? How come?

If you use the session mechanism the replication progress is synced with
the apply process, even if there are crashes. Crash recovery updates the
progress.  There's no such interlock with apply otherwise, and I don't
see how you can build one with reasonable effort.

Ah you mean because with wal_log=true the origin advance is in
different WAL record than commit? OK yeah you might be one transaction
behind then, true.

It actually means that we can not enforce database consistency. If we do
replorigin_advance  before commit and then crash happen, then we will
loose some changes.
If we call replorigin_advance after commit but crash happen before, then
some changes can be applied multiple times. For example we can insert
some record twice (if there are no unique constraints).
Look likes the only working scenario is to setup replication session for
each commit and use locking to prevent concurrent session setup for the
same slot by multiple process,  doesn't it?

You can do that, or you can move the tracking to the receiving process and spill the data to the disk (hurts IO obviously), or save the progress to table (also hurts IO), or write patch which solves this (no idea how though).

Also I concern about using sequential search for slot location
in replorigin_session_setup and many other functions - there is loop
through all max_replication_slots.
It seems to be not a problem when number of slots is less than 10. For
multimaster this assumption is true - even Oracle RAC rarely has
two-digit number of nodes.
But if we want to perform sharding and use logical replication for
providing redundancy, then number of nodes and slots can be essentially
I didn't think much about such configuration - may be it possible to
propose more efficient mechanism for replication in this case.

And each slot means connection with logical decoding attached to it so you don't really want to have thousands of those anyway. I think you'll hit other problems faster than loop over slots becomes problem if you plan to keep all of them active.

  Petr Jelinek                  http://www.2ndQuadrant.com/
  PostgreSQL Development, 24x7 Support, Training & Services

Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:

Reply via email to