Here is what I learned while working on an MSDN case with Microsoft in order to troubleshoot a serious degrade in performance that we faced following an upgrade from SQL Server 2000 to SQL Server 2005.
The distribution agent goes through the following steps to "move" a piece of data change from the distribution DB to a subscriber.
(1) querying the data change from the distribution DB
(2) moving the data change from distributor side to the subscriber side
(3) applying the data change to the subscriber DB
Step (1) is quite fast. Step (2) is slow in WAN. Step (3) is not fast. The distribution agent has two buffers.
In case of a push subscription, the distribution agent has two parallel tasks: (I) doing (1); (II) doing (2) and (3). At step (1), the data change is stored into a buffer of the distribution agent located at the distributor side.
In case of a pull subscription, the distribution agent has two parallel tasks: (I) doing (1) and (2); (II) doing (3). At step (2), the data change is stored into a buffer of the distribution agent located at the subscriber side.
We know that (2) and (3) are relatively slow. They are parallelized in a pull subscription, but serialized in a push subscription. So generally speaking, a pull subscription can perform better than a push subscription when it comes to WAN (Wide Area Network).
To leave a comment pls follow this link:
http://blogs.microsoft.co.il/blogs/yaniv_etrogi/archive/2009/03/03/pull-replication-vs-push-replication-part-ii.aspx
-Yaniv
============================================
