On Thursday 14 July 2005 08:36, Darcy Buskermolen wrote:
> Hello,
>
> I have implemented an experimental patch that avoids index maintainance
> during the SUBSCRIBE SET phase of replication.  This work is based apon
> theory and findings of Jan.   My initial testing shows that  subscribing a
> large heavily hit pgbench (scalling factor of 100 with 50 concurrent
> clients) database can subscribe set in the same amount of time as it takes
> to pg_dump | psql.
>
> Those of you who are interested can find the patch for 1.1 and CVS HEAD at
> http://www.dbitech.ca/slony/patches/subscription-perf-patch . I make no
> firm claim that, at this time, this patch will act as intended for
> everybody.  It would help us in getting this patch fully ironed out and
> applied to the main tree, if those of you who do try this patch would
> report your findings back to the list, wether it be yes it works or I got
> an error.


This patch has been update with a new version, this new version has some 
additional functionality, truncate is now used instead of delete to bypass 
the fkey limits that were there before.  A new slon.conf option has been 
introduced to allow the setting of maintenance_work_mem (vaccum_memory pre 
8.0) which is used on the subscriber nodes for reindexing/analyzing  the 
tables during the initial subscription phase.  This new option is named  
maintenance_work_mem and is only settable in slon.conf the default setting of 
0 means use sessions default value, otherwise we override it using the value 
specified (and reset it to the session default at the end of the 
subscription).  Any further feedback on this is most welcome. 

Outstanding TODO  related to this patch:
A) ample testing in as many odd setups as possible 
B) documentation
C) determine if the desired log shipping behavior is observed.



-- 
Darcy Buskermolen
Wavefire Technologies Corp.

http://www.wavefire.com
ph: 250.717.0200
fx: 250.763.1759