opendbteam workspace

opendbteam workspace
opendbteam workspace

Friday, April 29, 2016

#Postgresql #9.6 feature highlight: Multiple #synchronous #standbys

Among the features that are going to land with Postgres 9.6, here is a nice one that is going to make users in charge of high-availability structures quite happy, in short a lot:
commit: 989be0810dffd08b54e1caecec0677608211c339
author: Fujii Masao <>
date: Wed, 6 Apr 2016 17:18:25 +0900
Support multiple synchronous standby servers.

Previously synchronous replication offered only the ability to confirm
that all changes made by a transaction had been transferred to at most
one synchronous standby server.


Authors: Sawada Masahiko, Beena Emerson, Michael Paquier, Fujii Masao
Reviewed-By: Kyotaro Horiguchi, Amit Kapila, Robert Haas, Simon Riggs,
Amit Langote, Thomas Munro, Sameer Thakur, Suraj Kharage, Abhijit Menon-Sen,
Rajeev Rastogi
Behing this commit message wanted as incomplete for this post because it is really long, hides the possibility to set up multiple synchronous standby servers in a PostgreSQL cluster. Since 9.1, Postgres offers the possibility to have one, unique, synchronous standby, which guarantees that a transaction committed on a master node will never be lost on its synchronous standby node, assuming that synchronous_commit is set to ‘on’ (default), which ensures that when the transaction is committed on the master node the commit WAL record has been flushed to disk on the standby. Note that synchronous_commit can be set to some other values, feel free to have a look in the documentation and some explanation in last postregarding another new feature of 9.6 for more details.
With this feature one can have multiple standbys that guarantee that no transaction commit is lost on the way once committed on the master, giving the possibility to handle multiple failures in a cluster. Note that as the master needs to wait not for one, but for multiple standbys the confirmation that the commit WAL record has been flushed to disk before letting the client that the commit is finished localle, there is a performance penalty that gets higher the more synchronous standbys are used.
And actually, this feature is really great when combined with the new mode of synchronous_commit called remote_apply, because both features combined give the possibility to have a true read balancing among N nodes in a PostgreSQL cluster, and not only one. Some applications may want to give priority to this read balancing instead of cluster availability.
In order to use this feature, the grammar of synchronous_standby_names has been extended a bit with parenthesis separators, for example:
'N (standby1, standby2, ... standbyM)'
This means that N nodes are tracked as synchronous among the set of M standbys defined in the list. Note as well the following:
  • If N > M, all the nodes are considered as synchronous. And actually be careful not to do that, because this would make the code wait for nodes that do not exist. Note that when this happens the server generates a WARNING in its logs.
  • If N < M, the first N nodes listed and currently connected to a master node are considered as synchronous, and the rest are potential candidates to be synchronous.
  • If N = M, all the nodes are considered as synchronous
  • N = 1 is equivalent to the pre-9.5 grammar, the case where no parenthesis separators are used in the string value.
Now let’s see the feature in action with one master and four standbys, switching them so as 2 synchronous standbys are set with one standby that could potentially become synchronous if one of the existing synchronous nodes goes missing:
=# ALTER SYSTEM SET synchronous_standby_names = '2(node_5433, node_5434, node_5435)';
=# SELECT pg_reload_conf();
(1 row)
=# SELECT application_name, sync_priority, sync_state
   FROM pg_stat_replication ORDER BY application_name;
 application_name | sync_priority | sync_state
 node_5433        |             1 | sync
 node_5434        |             2 | sync
 node_5435        |             3 | potential
 node_5436        |             0 | async
(4 rows)
=# SHOW synchronous_commit;
(1 row)
And done. With synchronous_commit set to ‘on’, transaction commits are guaranteed to not be lost on those two nodes, improving the whole system availability.

Wednesday, April 20, 2016

Spring is here! Time for another #Benchmark using #PostgreSQL with differents types of disks...

Ever wondered how SSDs compare with HDDs?
And whether PostgreSQL performance improve with SSD?
Are tablespaces necessary?

Well, search no more and read this article by Tomas Vondra: