Jack, I think the issue is that the ping which is used to determine whether or not the server is live returns a seemingly false positive back to the load balancer (and indirectly the client) that this server is ready to go when in fact it's not. Reading this page ( http://wiki.apache.org/solr/SolrConfigXml), it does seem to be documented to do this but it may not be fully stressed to hide your Solr behind a load balancer. I am more than happy to write up a post that, in my opinion at least, stresses some best practices on the use of Solr based on my experience if others find this useful.
What seems odd here is that the ping is a query so maybe the ping query in the solrconfig (for Aaron and others having this) should be configured to hit the handler that is used by the front end app so that while that handler is warming up the ping query will be blocked. Of course using the load balancer means that the app layer knows nothing about servers in and out of rotation. Cheers! Amit On Sun, Nov 11, 2012 at 8:05 AM, Jack Krupansky <j...@basetechnology.com>wrote: > Is the issue here that the Solr node is continuously "live" with the load > balancer so that the moment during startup that Solr can respond to > anything, the load balancer will be sending it traffic and that this can > occur while Solr is still warming up? > > First, shouldn't we be encouraging people to have an app layer between > Solr and the outside world? If so, the app layer should simply not respond > to traffic until the app layer can verified that Solr has stabilized. If > not, then maybe we do need to suggest a change to Solr so that the > developer can control exactly when Solr becomes "live" and responsive to > incoming traffic. > > At a minimum, we should document when that moment is today in terms of an > explicit contract. It sounds like the problem is that the contract is > either nonexistent, vague, ambiguous, non-deterministic, or whatever. > > -- Jack Krupansky > > -----Original Message----- From: Amit Nithian > Sent: Saturday, November 10, 2012 4:24 PM > To: solr-user@lucene.apache.org > Subject: Re: Preventing accepting queries while custom QueryComponent > starts up? > > > Yeah that's what I was suggesting in my response too. I don't think your > load balancer should be doing this but whatever script does the release > (restarting the container) should do this so that when the ping is enabled > the warming has finished. > > > On Sat, Nov 10, 2012 at 3:33 PM, Erick Erickson <erickerick...@gmail.com>* > *wrote: > > Hmmm, rather than hit the ping query, why not just send in a real query >> and >> only let the queued ones through after the response? >> >> Just a random thought >> Erick >> >> >> On Sat, Nov 10, 2012 at 2:53 PM, Amit Nithian <anith...@gmail.com> wrote: >> >> > Yes but the problem is that if user facing queries are hitting a server >> > that is warming up and isn't being serviced quickly, then you could >> > potentially bring down your site if all the front end threads are > >> blocked >> > on Solr queries b/c those queries are waiting (presumably at the >> container >> > level since the filter hasn't finished its init() sequence) for the >> warming >> > to complete (this is especially notorious when your front end is rails). >> > This is why your ping to enable/disable a server from the load balancer >> has >> > to be accurate with regards to whether or not a server is truly ready > >> and >> > warm. >> > >> > I think what I am gathering from this discussion is that the server is >> > warming up, the ping is going through and tells the load balancer this >> > server is ready, user queries are hitting this server and are queued >> > waiting for the firstSearcher to finish (say these initial user queries >> are >> > to respond in 500-1000ms) that's terrible for performance. >> > >> > Alternatively, if you have a bunch of servers behind a load balancer, > >> you >> > want this one server (or block of servers depending on your deployment) >> to >> > be reasonably sure that user queries will return in a decent time >> (whatever >> > you define decent to be) hence why this matters. >> > >> > Let me know if I am missing anything. >> > >> > Thanks >> > Amit >> > >> > >> > On Sat, Nov 10, 2012 at 10:03 AM, Erick Erickson < >> erickerick...@gmail.com >> > >wrote: >> > >> > > Why does it matter? The whole idea of firstSearcher queries is to warm >> up >> > > your system as fast as possible. The theory is that upon restarting > >> > the >> > > server, let's bet this stuff going immediately... They were never >> > intended >> > > (as far as I know) to complete before any queries were handled. As an >> > > aside, I'm not quite sure I understand why pings during the warmup are >> a >> > > problem. >> > > >> > > But anyway. firstSearcher is particularly relevant because the >> > > autowarmCount settings on your caches are irrelevant when starting the >> > > server, there's no history to autowarm.... >> > > >> > > But, there's no good reason _not_ to let queries through while >> > > firstSearcher is doing it's tricks, they just get into the queue and >> are >> > > served as quickly as they may. That might be some time since, as you >> say, >> > > they may not get serviced until the expensive parts get filled. But I >> > don't >> > > think having them be serviced is doing any harm. >> > > >> > > Now, newSearcher and autowarming of the caches is a completely >> different >> > > beast since having the old searchers continue serving requests until >> the >> > > warmups _does_ directly impact the user, they don't see random > > >> slowness >> > > because a searcher is being opened. >> > > >> > > So I guess my real question is whether you're seeing a measurable >> problem >> > > or if this is a red herring.... >> > > >> > > FWIW, >> > > Erick >> > > >> > > >> > > On Thu, Nov 8, 2012 at 2:54 PM, Aaron Daubman <daub...@gmail.com> >> wrote: >> > > >> > > > Greetings, >> > > > >> > > > I have several custom QueryComponents that have high one-time > > > >> startup >> > > costs >> > > > (hashing things in the index, caching things from a RDBMS, etc...) >> > > > >> > > > Is there a way to prevent solr from accepting connections before all >> > > > QueryComponents are "ready"? >> > > > >> > > > Especially, since many of our instance are load-balanced (and >> > > > added-in/removed automatically based on admin/ping responses) >> > preventing >> > > > ping from answering prior to all custom QueryComponents being ready >> > would >> > > > be ideal... >> > > > >> > > > Thanks, >> > > > Aaron >> > > > >> > > >> > >> >> >