hello shawn,

thanks for the reply.

ok - i did some testing and yes you are correct.  

autocommit is doing the "commit" work in chunks. yes - the slaves are also
going to having everything to nothing, then slowly building back up again,
lagging behind the master.

... and yes - this is probably not what we need - as far as a replication
strategy for the slaves.

you said, you don't use autocommit.  if so - then why don't you use / like
autocommit?

since we have not done this here - there is no established reference point,
from an operations perspective.

i am looking to formulate some sort of operation strategy, so ANY ideas or
input is really welcome.



it seems to me that we have to account for two operational strategies - 

the first operational mode is a "daily" append to the solr core after the
database tables have been updated.  this can probably be done with a simple
delta import.  i would think that autocommit could remain on for the master
and replication could also be left on so the slaves picked up the changes
ASAP.  this seems like the mode that we would / should be in most of the
time.


the second operational mode would be a "build from scratch" mode, where
changes in the schema necessitated a full re-index of the data.  given that
our site (powered by solr) must be up all of the time, and that our full
index time on the master (for the moment) is hovering somewhere around 16
hours - it makes sense that some sort of parallel path - with a cut-over,
must be used.

in this situation is it possible to have the indexing process going on in
the background - then have one commit at the end - then turn replication on
for the slaves?

are there disadvantages to this approach?

also - i really like your suggestion of a "build core" and "live core".  is
this approach you use?

thank you for all of the great input




then 


--
View this message in context: 
http://lucene.472066.n3.nabble.com/should-slave-replication-be-turned-off-on-during-master-clean-and-re-index-tp3945531p3952904.html
Sent from the Solr - User mailing list archive at Nabble.com.

Reply via email to