If using the embedded for three nodes (for example) solr you will get one of two things out of the box:
- Start zk on every one of 3 nodes, get 3 independent clusters that know nothing of each other - Start zk on one "key" node, point the other two nodes at the key node, get one cluster that shards can be distributed across, but if your key node goes down all goes down These can be fine for small, simple use cases that are not mission critical, and/or where queries and load are known to be well behaved and risk is more acceptable than. I've not tried to convince the embedded zk's to form a zk cluster and I don't think that that process is well documented, but probably one could achieve that manually with a little research/work, probably mostly tracking down where the zk config is landing and manually editing it and then ensuring somehow that those changes are not easily lost, but it does not happen OOTB. What I "brainstormed" in the prior email amounts to figuring that out, documenting it and writing code that can make it happen automatically. (some need for discovery/management too). I think the biggest thing is for folks to have clarity on what risks they are taking, and if they have decided to accept the risk with full appreciation for it, that's fine. Important to also realize that zero risk doesn't exist and striving for it may not make economic sense in all cases. Being irrationally scared of risk is also a pitfall. On Thu, Aug 12, 2021 at 7:09 AM Uwe Schindler <u...@thetaphi.de> wrote: > I am fully happy with something that works out of box. > > > > The main problems I see with many customers is not only the complexity of > setup, but also that you need to install a separate Zookeeper ensemble. > When you tell them: Come on, use the one provided by a solr node and you > are fine: “no this is not allowed, see doc xy”. > > > > So let us please simplify the recommendations: If you have one or 2 or > three nodes in standalone node, it is perfectly fine to use embedded > zookeeper. We should not overreact here. A user who used Master/Slave > replication is also not fully fault tolerant. > > > > I’d change the documentation to say something: “If you want to scale, use > a separate zookeeper ensemble with a minimum of three nodes. But for simple > setups just relying on the good old master/slave replication (not the > default solr one that distributes indexing), it is perfectly fine to use > embedded zookeeper (on the “master” node that holds the main index). This > setup is then not really different from classical master/slave replication. > > > > As said before, I am not against Solr cloud, but lets keep it simple for > people that want to keep it simple. I am also fine to start a single node > cluster with zookeeper, but this should be the embedded one (just as > datastore for the fake cluster). And no warnings should be printed. Maybe > as soon as you add too many nodes, print some warning “now it is time to > setup a separate zookeeper ensemble”. But, please not for 2 nodes > (master/slave). > > > > Also where is the problem in spawning an embedded zookeeper in every node > by default? Why does it need to be separated? > > > > Uwe > > > > ----- > > Uwe Schindler > > Achterdiek 19, D-28357 Bremen > > https://www.thetaphi.de > > eMail: u...@thetaphi.de > > > > *From:* Jan Høydahl <jan....@cominvent.com> > *Sent:* Wednesday, August 11, 2021 4:27 PM > *To:* dev@solr.apache.org > *Subject:* Re: SolrCloud Alone: Deprecate Standalone Mode > > > > However, we tell people not to use the embedded ZK in production, so I’m > curious if that’s only because it’s a single-node ZK or if there is > something else about the way we’ve embedded it that we would need to change? > > > > As I recall there are several reasons. First, our embedded ZK was kind of > a hack with some forked code etc. Second, it is not designed to be fault > tolerant even if you start three solr nodes this way we cannot form a > quorum. And perhaps third, ZK has not been officially supported on > Windows.. However, I believe this is all solvable if we want to day. Not > saying it is easy though :) > > > > Jan > > > > 11. aug. 2021 kl. 16:17 skrev Cassandra Targett <casstarg...@gmail.com>: > > > > So basically the proposal would be that we use the embedded ZK to > automatically create a quorum via multiple nodes. That’s an interesting > idea. > > However, we tell people not to use the embedded ZK in production, so I’m > curious if that’s only because it’s a single-node ZK or if there is > something else about the way we’ve embedded it that we would need to change? > > I was also under the impression that beyond the complexities of ZK there > are still use cases that SolrCloud does not adequately support, even with > the addition of TLOG and PULL replicas. Does anyone have any examples of > those? > > I’d also like to remind folks to please not use the terminology > “master/slave”, we removed it from the code and documentation because it’s > not inclusive for our community. > > Similarly, “standalone” has always been rather imprecise - it’s not > “standalone”, it’s a cluster but without ZK and other automation sugar. In > the Ref Guide we’ve settled on “user-managed”. It sounds pedantic but it > matters because we should be really clear about what we’re talking about - > deprecating and removing the ability for a single-node Solr installation? > Only the mode of a non-ZK cluster? Both? > > On Aug 11, 2021, 6:39 AM -0500, Eric Pugh <ep...@opensourceconnections.com>, > wrote: > > For small setups I’ve used a single ZK and a single Solr node very > successfully, the operational benefits of all the SolrCloud API’s has been > fantastic. > > > > I’ve always thought that us having ZooKeeper as this “front and center” > requirement for SolrCloud was always a weird decision that would put off a > lot of folks. We don’t beat our potential users over the head with the > fact we use Jetty for example. It’s just part of the stack. > > > > The flow that Gus proposed should have been added to SolrCloud a long time > ago, how much easier would it have made all our lives! The entire > existence of ZooKeeper should be behind APIs and be an abstraction. We > should do this regardless of if deprecated standalone! > > > > Uwe, if we had what Gus proposed, but eliminate zk, would that map much > more to what you wanted? Here is my attempt at retelling the story that > Gus told, but to meet the goals of folks who might want to move to ES for > ease: > > > > A) Start Node 1. > B) Start Node 2 telling it that Node 1 exists. node 2 comes up, joins > network and messages “at risk for split brain”. > C) Start Node 3 telling it that node 1 exists. node 1, node 2, node 3 all > under the covers are sharing state via ZK and messages “no risk for split > brain" > D) Node 4 - like node 2 but since we have optimum quorum doesn’t add to ZK > (under covers, hidden from user). > E) Node 5 - like node 3, but since we have optimum quorum doesn’t add to > ZK (under covers, hidden from user). > > > > > > > > > > > > On Aug 11, 2021, at 7:15 AM, Uwe Schindler <u...@thetaphi.de> wrote: > > > > Hi, > > > > most of my customers prefer standalone mode and manual replication. A lot > of setups, especially in Germany, are very > > > > Solr Cloud is only interesting to large customers that want to scale > hugely. But from what I have seen, most of those have moved to > Elasticsearch or Opensearch (see below). The biggest issue is always the > stupidness of having to maintain a separate Zookeeper cloud, which adds > more hardware/VMs to the game and makes the thing more complex. If you want > to maintain up to 4 or 6 Solr nodes with one index and a few shards, the > overhead by Zookeeper (you need 3 of them) is – sorry to say – > unmaintainable. With Elasticsearch it’s easy to setup. No dedicated > cloud/standalone mode. You just start a single node and test it. If it > works fine, you start additional nodes to form a cloud. Plain simple. > Config files are easy to handle, you need no ip addresses hardcoded into > Zookeeper nodes, it just works. If you don’t want to make people move to > Elasticsearch/Opensearch, make them happy with their fully controllable > local master/slave mode. > > > > So my strong -1 to make cloud mode the default and deprecate standalone > mode. Unless both is the same and works without a separate zookeeper > cluster, I won’t change my vote. > > > > Uwe > > > > ----- > > Uwe Schindler > > Achterdiek 19, D-28357 Bremen > > https://www.thetaphi.de > > eMail: u...@thetaphi.de > > > > *From:* Gus Heck <gus.h...@gmail.com> > *Sent:* Tuesday, August 10, 2021 8:34 PM > *To:* dev@solr.apache.org > *Subject:* Re: SolrCloud Alone: Deprecate Standalone Mode > > > > Or to keep things fast without retaining all the checks, one could provide > slow/fast modes for test, fast requiring a local zookeeper external to the > tests, with the tests properly namespacing themselves... that does imply > reworking some tests. > > > > Now that I say the above, it would be interesting if the some of the tests > could (also optionally) properly isolate themselves within an externally > running solr (probably started via cloud.sh with the latest edits. ... > develop, cloud.sh, test manually, run tests against same I expect that > there are still tests for which that makes no sense of course. This is > probably a crazier idea than using an external zookeeper however, where > zkChroot should be sufficient to isolate things I think... > > > > -Gus > > > > On Tue, Aug 10, 2021 at 2:22 PM David Smiley <dsmi...@apache.org> wrote: > > Good call-out on perceived complexity due to running 3 ZK nodes. For many > small installations, honestly Solr's embedded ZK is fine. Also, again for > small installations, running ZK alongside Solr (same hardware) is fine. We > shouldn't needlessly shame users away from doing these things as if it's > irresponsible. There's a spectrum of demands on Solr from low to high. > Anyway, I suspect it's increasingly moot with more Docker & Kubernetes > being used to reduce the hassles of deploying any service (be it Solr or > whatever). This will only increase going forward. > > > > Even if ZK becomes the only mode, I expect many checks in our codebase > that conditionally check for ZK to remain. We want tests that don't care > about SolrCloud mode to be fast, and that means not running unnecessary > things like ZooKeeper. > > > ~ David Smiley > > Apache Lucene/Solr Search Developer > > http://www.linkedin.com/in/davidwsmiley > > > > > > On Tue, Aug 10, 2021 at 12:23 PM Gus Heck <gus.h...@gmail.com> wrote: > > I've met several clients who really didn't want to manage zookeeper as an > additional service (I've talked some into it anyway, but it was clearly a > key reason they hadn't started/gone cloud). I think it would be far more > palatable if it's all "part of solr", doesn't require plumbing the docs of > some other project entirely, and requires neither requisitioning additional > hardware nor service scripts, monitoring, support that isn't "solr" > support... etc... then I think that alleviates some of the pain that folks > in small sub-sections of moderate to large orgs feel at the idea of using > cloud. These folks face long procurement cycles and disaster/recovery plans > etc, despite often having team sizes under 20... or face having to educate > large IT departments into handling deployments when they themselves are new > (of course that's how some of them wind up hiring folks like me... but > that's a barrier too since that has to be approved too). Also I've met > folks who didn't understand that it was possible to have a 1 node "cluster" > with zk on the same machine, and had the impression that 5 boxes (2 solr > and 3 zk) were absolutely required to run cloud. Which it is of course for > high availability with no SPOF, but it is not required if you don't need > high availability. > > > > I think to sunset "user managed" we need to figure out how to self manage > embedded zookeepers, most particularly setup for smaller orgs or lower > traffic installs should look like: > > > > A) Start Node 1 with zk embedded ... if you only need one node, don't want > high availability etc, done. > > B) Start Node 2 telling it the zk url for node 1. node 2 comes up, offers > to participate in zk, but does not because that would make an even number > > C) Start Node 3 telling it the zk url for node 1. node 1 (node 2 hasn't > started zk) node 3 offers to participate in zk, and now with 2 offers > pending, both 2 and 3, get up to date on the current state and th join, now > the embedded zk cluster is 3 nodes, not one, and no SPOF... as they grow... > > D) Node 4 - like node 2 but can use zk url of any of 1,2,3 > > E) Node 5 - like node 3, but can use zk url of any 1,2,3 > > > > Obviously, features for users to set a cap the size of zk clusters, don't > need 49 nodes on 50 servers... , ensure they put their data in a convenient > place that is well documented, document how to secure the inter-node > connections, clarity in the admin UI of what nodes have zk etc. > > > > For this embedded zk use case we should document whatever the user needs > to know so they don't have to sort through docs at an entirely different > project not necessarily focused on the things solr users need. > > > > Certainly we would still advocate for a separate zk cluster for better > performance/stability. In essence a supported mode with known > limitations... True we have to support all THAT code instead, but the > available feature set becomes consistent and a bazillion checks to see if > we have zkStateReader (or some other sentinel for cloud mode) can > disappear, so probably a net gain etc. > > > > On the flip side I"ve also had the thought that cluster state management > should be pluggable such that if a better tool than zk, or merely an > "already installed" tool is available solr could use it. Without careful > thought everything I just said could take us in the opposite direction > > > > Maybe running zk embedded is "Solr Fog" mode :) > > > > On Mon, Aug 9, 2021 at 2:55 PM Houston Putman <houstonput...@gmail.com> > wrote: > > I agree with David that the first step would be to make SolrCloud the > default mode. > > I made a dev list thread about this a few months ago, but I think I failed > to respond at some point. > > I will get back on that and address the > > > > I also really like Mike's idea that we enable very similar use cases with > embedded Zookeeper's, > > if at all possible, to make the transition easy for users who want to stay > on the user-manager mode. > > > > Marcus, I think it would be a great idea to fix up the documentation to > make SolrCloud the first and most prominent mode advertised. > > Never saw your original PR, but would love to give it a look if you > resuscitate it at some point. > > > > - Houston > > > > On Mon, Aug 9, 2021 at 2:48 PM David Smiley <dsmi...@apache.org> wrote: > > Given that SolrCloud is not even the default mode, I think it is premature > to deprecate standalone mode. Let's do this first and maybe consider > deprecating standalone after some time? > > > ~ David Smiley > > Apache Lucene/Solr Search Developer > > http://www.linkedin.com/in/davidwsmiley > > > > > > On Mon, Aug 9, 2021 at 1:58 PM Mike Drob <md...@mdrob.com> wrote: > > Could we simulate user managed replication with an embedded zookeeper > on the primary and pull replicas on the followers? > > On Mon, Aug 9, 2021 at 12:56 PM Jason Gerlowski <gerlowsk...@gmail.com> > wrote: > > > > Hey Marcus, > > > > The places I've worked in the past have all used SolrCloud primarily > > so I can't speak to any specifics, but my impression from reading > > user-list traffic is that a sizable chunk of Solr's user base prefers > > "User-Managed" mode (formerly called "standalone"). Some because they > > don't want to manage a ZooKeeper cluster. Some because the > > replication model in 'user-managed' fits their needs better. Some I > > imagine just haven't bothered to update in many years. > > > > I'm absolutely sympathetic to efforts to streamline development and > > reduce collective debt, but it might be tough to displace such a big > > chunk of users. I'm curious what others think though. Maybe the > > proportion of 'user-managed' users out there is smaller than I > > imagine. > > > > Jason > > > > On Fri, Aug 6, 2021 at 11:59 PM Marcus Eagan <marcusea...@gmail.com> > wrote: > > > > > > Hello again, > > > > > > Has the time come for us to reduce scope to move faster and with more > focus? Even for those not in the cloud, SolrCloud has been the undisputed > performance and usability champ since version 8.0. In version 9.0, I'd like > to propose that the deciders in the community deprecate standalone mode in > favor of SolrCloud. > > > > > > There are a few drivers: > > > > > > We only need to support changes that impact SolrCloud going forward. I > know that this is hard to stomach. But by the time Solr reaches version > 10.0, everyone should have migrated to SolrCloud as there is little reason > to continue to run standalone. > > > The new features keep coming to SolrCloud, but not to standalone. You > can see in a few ways how I embarrassingly discovered this late one night > while trying out a PR. If not careful, users can accidentally start Solr in > standalone mode. Think of all the features that they will see documented > but not in their environment. What a confusing user experience? > > > Last but certainly not least, the number of contributors to the > project, and the velocity of those contributions has dropped. . It does not > have to be that way, though. Two ways are for the community to observe our > push for modernization and improved user experience. Simply eliminating the > need to include the -c flag in the start command would be a huge win for > many engineers.We should make life easier for our users as much as the > maintainers here. We can strive to make the upgrade process from 9 to 10 > very simple. > > > > > > I tried to make one step in this direction last year by re-ordering > the README to show the Solr Cloud command before the standalone command. I > believe that patch died on the vine, but I would be excited to revive it to > document this effort when the time is appropriate. > > > > > > Reason not to do it: > > > > > > Some large company out there might view this move as introducing > risk. I view the risk here as negligible but I welcome any perspective > there. > > > Some things I inevitably don't know. > > > > > > What do you all think? > > > > > > Thank you all for your voluntary contributions, > > > -- > > > Marcus Eagan > > > > > > > --------------------------------------------------------------------- > > To unsubscribe, e-mail: dev-unsubscr...@solr.apache.org > > For additional commands, e-mail: dev-h...@solr.apache.org > > > > --------------------------------------------------------------------- > To unsubscribe, e-mail: dev-unsubscr...@solr.apache.org > For additional commands, e-mail: dev-h...@solr.apache.org > > > > > -- > > http://www.needhamsoftware.com (work) > > http://www.the111shift.com (play) > > > > > -- > > http://www.needhamsoftware.com (work) > > http://www.the111shift.com (play) > > > > _______________________ > > *Eric Pugh **| *Founder & CEO | OpenSource Connections, LLC | > 434.466.1467 | http://www.opensourceconnections.com | My Free/Busy > <http://tinyurl.com/eric-cal> > > Co-Author: Apache Solr Enterprise Search Server, 3rd Ed > <https://www.packtpub.com/big-data-and-business-intelligence/apache-solr-enterprise-search-server-third-edition-raw> > > This e-mail and all contents, including attachments, is considered to be > Company Confidential unless explicitly stated otherwise, regardless > of whether attachments are marked as such. > > > > > -- http://www.needhamsoftware.com (work) http://www.the111shift.com (play)
