lucene-solr-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Michael Della Bitta <michael.della.bi...@appinions.com>
Subject Re: Poll: SolrCloud vs. Master-Slave usage
Date Fri, 01 Mar 2013 13:37:43 GMT
Amit,

NRT is not possible in a master-slave setup because of the necessity
of a hard commit and replication, both of which add considerable
delay.

Solr Cloud sends each document for a given shard to each node hosting
that shard, so there's no need for the hard commit and replication for
visibility.

You could conceivably get NRT on a single node without Solr Cloud, but
there would be no redundancy.

Michael Della Bitta

------------------------------------------------
Appinions
18 East 41st Street, 2nd Floor
New York, NY 10017-6271

www.appinions.com

Where Influence Isn’t a Game


On Fri, Mar 1, 2013 at 1:22 AM, Amit Nithian <anithian@gmail.com> wrote:
> Erick,
>
> Well put and thanks for the clarification. One question:
> "And if you need NRT, you just can't get it with traditional M/S setups."
> ==> Can you explain how that works with SolrCloud?
>
> I agree with what you said too because there was an article or discussion I
> read that said having high-availability masters requires some fairly
> complicated setups and I guess I am under-estimating how
> expensive/complicated our setup is relative to what you can get out of the
> box with SolrCloud.
>
> Thanks!
> Amit
>
>
> On Thu, Feb 28, 2013 at 6:29 PM, Erick Erickson <erickerickson@gmail.com>wrote:
>
>> Amit:
>>
>> It's a balancing act. If I was starting fresh, even with one shard, I'd
>> probably use SolrCloud rather than deal with the issues around the "how do
>> I recover if my master goes down" question. Additionally, SolrCloud allows
>> one to monitor the health of the entire system by monitoring the state
>> information kept in Zookeeper rather than build a monitoring system that
>> understands the changing topology of your network.
>>
>> And if you need NRT, you just can't get it with traditional M/S setups.
>>
>> In a mature production system where all the operational issues are figured
>> out and you don't need NRT, it's easier just to plop 4.x in traditional M/S
>> setups and not go to SolrCloud. And you're right, you have to understand
>> Zookeeper which isn't all that difficult, but is another moving part and
>> I'm a big fan of keeping the number of moving parts down if possible.
>>
>> It's not a one-size-fits-all situation. From what you've described, I can't
>> say there's a compelling reason to do the SolrCloud thing. If you find
>> yourself spending lots of time building monitoring or High
>> Availability/Disaster Recovery tools, then you might find the cost/benefit
>> analysis changing.
>>
>> Personally, I think it's ironic that the memory improvements that came
>> along _with_ SolrCloud make it less necessary to shard. Which means that
>> traditional M/S setups will suit more people longer <G>....
>>
>> Best
>> Erick
>>
>>
>> On Thu, Feb 28, 2013 at 8:22 PM, Amit Nithian <anithian@gmail.com> wrote:
>>
>> > I don't know a ton about SolrCloud but for our setup and my limited
>> > understanding of it is that you start to bleed operational and
>> > non-operational aspects together which I am not comfortable doing (i.e.
>> > software load balancing). Also adding ZooKeeper to the mix is yet another
>> > thing to install, setup, monitor, maintain etc which doesn't add any
>> value
>> > above and beyond what we have setup already.
>> >
>> > For example, we have a hardware load balancer that can do the actual load
>> > balancing of requests among the slaves and taking slaves in and out of
>> > rotation either on demand or if it's down. We've placed a virtual IP on
>> top
>> > of our multiple masters so that we have redundancy there. While we have
>> > multiple cores, the data volume is large enough to fit on one node so we
>> > aren't at the data volume necessary for sharding our indices. I suspect
>> > that if we had a sufficiently large dataset that couldn't fit on one box
>> > SolrCloud is perfect but when you can fit on one box, why add more
>> > complexity?
>> >
>> > Please correct me if I'm wrong for I'd like to better understand this!
>> >
>> >
>> >
>> >
>> > On Thu, Feb 28, 2013 at 12:53 AM, rulinma <rulinma@gmail.com> wrote:
>> >
>> > > I am doing research on SolrCloud.
>> > >
>> > >
>> > >
>> > > --
>> > > View this message in context:
>> > >
>> >
>> http://lucene.472066.n3.nabble.com/Poll-SolrCloud-vs-Master-Slave-usage-tp4042931p4043582.html
>> > > Sent from the Solr - User mailing list archive at Nabble.com.
>> > >
>> >
>>

Mime
View raw message