Sql-server – SQL Server 2008 Replication – Fixed Overhead of Replication

Running on a VM is a good idea as it's hard to define resource requirements but in most cases distributors are quite lightweight. The main factors are number of subscribers, transactions count and the number of articles in replication.

Some tips to help:

• Only replicate what you need to, this includes SPs, views etc.
• Using pull will reduce CPU and memory requirements on the distributor.
  • This is harder to implement in Web edition but possible with scripts.
  • If you have a lot of push agents running, you should allow more memory outside of SQL.
  • Consider increasing the desktop heap size if using push and have a lot of agents. (this is a registry change)
• Choose the right type of replication.
  • Snapshot is light on distributors as it doesn’t have the overhead of the log reader or clean-up jobs but heavier on the network and the subscriber when pushing out.
• Enterprise edition doesn’t provide any real performance benefits on the distributor.
  • I hoped that merry-go-round reads would but no, it’s mostly in CPU when problems occur.
• You may find it helpful to reserve some cores for the network as large distributors can get overwhelmed during imports or clean-up jobs and struggle to get the data out.
  • I’m talking 1000 subscribers and 100,000 articles (not my idea).
• Throwing CPU at the distributor rarely fixes the high CPU issue. You need to treat the distribution databases like any other database and tune it (I believe Denny Cherry has an article which recommends some indexes).
• CPU can also be caused by the locking and deadlocking during the clean-up job or large transactions failing to be delivered causing the subscriber agents to scan the entire table. You may find that reducing the frequency of the clean-up job actually help overall throughput but could impact latency.

In a nutshell, if replication is configured correctly, updates are batched properly (don’t update millions of rows in one transaction) and the subscribers are up to date then it’s happy.

Below script will help you find out how many undistributed commands are there. Also, do you see any errors in sql server or windows logs ? Is it just slow or any failures reported ?

SELECT ( CASE 
           WHEN mdh.runstatus = '1' THEN 'Start - ' + CAST( 
                                         mdh.runstatus AS VARCHAR) 
           WHEN mdh.runstatus = '2' THEN 'Succeed - ' + CAST( 
                                         mdh.runstatus AS VARCHAR) 
           WHEN mdh.runstatus = '3' THEN 'InProgress - ' + CAST( 
                                         mdh.runstatus AS VARCHAR) 
           WHEN mdh.runstatus = '4' THEN 'Idle - ' + CAST(mdh.runstatus AS 
                                                          VARCHAR) 
           WHEN mdh.runstatus = '5' THEN 'Retry - ' + CAST( 
                                         mdh.runstatus AS VARCHAR) 
           WHEN mdh.runstatus = '6' THEN 'Fail - ' + CAST(mdh.runstatus AS 
                                                          VARCHAR) 
           ELSE CAST(mdh.runstatus AS VARCHAR) 
         END )                                   [Run Status], 
       mda.subscriber_db                         [Subscriber DB], 
       mda.publication                           [PUB Name], 
       RIGHT(LEFT(mda.name, Len(mda.name) - ( Len(mda.id) + 1 )), 
       Len(LEFT(mda.name, Len(mda.name) - ( Len(mda.id) + 1 ))) - ( 
       10 + Len(mda.publisher_db) + ( CASE 
                                          WHEN mda.publisher_db = 'ALL' THEN 1 
                                          ELSE Len(mda.publication) + 2 
                                        END ) )) [SUBSCRIBER], 
       CONVERT(VARCHAR(25), mdh.[time])          [LastSynchronized], 
       und.undelivcmdsindistdb                   [UndistCom], 
       mdh.comments                              [Comments], 
       'select * from distribution.dbo.msrepl_errors (nolock) where id = ' + 
       CAST(mdh.error_id AS VARCHAR(8))          [Query More Info], 
       mdh.xact_seqno                            [SEQ_NO], 
       ( CASE 
           WHEN mda.subscription_type = '0' THEN 'Push' 
           WHEN mda.subscription_type = '1' THEN 'Pull' 
           WHEN mda.subscription_type = '2' THEN 'Anonymous' 
           ELSE CAST(mda.subscription_type AS VARCHAR) 
         END )                                   [SUB Type], 
       mda.publisher_db + ' - ' + CAST(mda.publisher_database_id AS VARCHAR) 
                                                 [Publisher DB], 
       mda.name 
       [Pub - DB - Publication - SUB - AgentID] 
FROM   distribution.dbo.msdistribution_agents mda 
       LEFT JOIN distribution.dbo.msdistribution_history mdh 
         ON mdh.agent_id = mda.id 
       JOIN (SELECT s.agent_id, 
                    maxagentvalue.[time], 
                    SUM(CASE 
                          WHEN xact_seqno > maxagentvalue.maxseq THEN 1 
                          ELSE 0 
                        END) AS undelivcmdsindistdb 
             FROM   distribution.dbo.msrepl_commands t (nolock) 
                    JOIN distribution.dbo.mssubscriptions AS s (nolock) 
                      ON ( t.article_id = s.article_id 
                           AND t.publisher_database_id = s.publisher_database_id 
                         ) 
                    JOIN (SELECT hist.agent_id, 
                                 MAX(hist.[time]) AS [time], 
                                 h.maxseq 
                          FROM   distribution.dbo.msdistribution_history hist ( 
                                 nolock) 
                                 JOIN (SELECT agent_id, 
                                              Isnull(MAX(xact_seqno), 0x0) AS 
                                              maxseq 
                                       FROM 
                                 distribution.dbo.msdistribution_history ( 
                                 nolock) 
                                       GROUP  BY agent_id) AS h 
                                   ON ( hist.agent_id = h.agent_id 
                                        AND h.maxseq = hist.xact_seqno ) 
                          GROUP  BY hist.agent_id, 
                                    h.maxseq) AS maxagentvalue 
                      ON maxagentvalue.agent_id = s.agent_id 
             GROUP  BY s.agent_id, 
                       maxagentvalue.[time]) und 
         ON mda.id = und.agent_id 
            AND und.[time] = mdh.[time]