On Wed, Jul 6, 2016 at 2:30 AM, Lorenzo Fundaró < lorenzo.fund...@dawandamail.com> wrote:
> On 6 July 2016 at 00:00, Tomás Fernández Löbbe <tomasflo...@gmail.com> > wrote: > > > The leader will do the replication before responding to the client, so > lets > > say the leader gets to update it's local copy, but it's terminated before > > sending the request to the replicas, the client should get either an HTTP > > 500 or no http response. From the client code you can take action (log, > > retry, etc). > > > > If this true then whenever I ask for min_rf having three nodes (1 leader + > 2 replicas) > I should get rf = 3, but in reality i don't. > > > > The "min_rf" is useful for the case where replicas may be down or not > > accessible. Again, you can use this for retrying or take any necessary > > action on the client side if the desired rf is not achieved. > > > > > I think both paragraphs are contradictory. If the leader does the > replication before responding to the client, then > why is there a need to use the min_rf ? I don;t think is true that you get > a 200 when the update has been passed to all replicas. > The reason why "min_rf" is there is because: * If there are no replicas at the time of the request (e.g. if replicas are unreachable and disconnected from ZK) * Replicas could fail to ACK the update request from the leader, in that case the leader will mark them as unhealthy but would HTTP 200 to the client. So, it could happen that you think your data is being replicated to 3 replicas, but 2 of them are currently out of service, this means that your doc is in a single host, and if that one dies, then you lose that data. In order to prevent this, you can ask Solr to tell you how many replicas succeeded that update request. You can read more about this in https://issues.apache.org/jira/browse/SOLR-5468 > > The thing is that, when you have persistent storage yo shouldn't worry > about this because you know when the node comes back > the rest of the index will be sync, the problem is when you don't have > persistent storage. For my particular case I have to be extra careful and > always > make sure that all my replicas have all the data I sent. > > In any case you should assume that storage on a host can be completely lost, no mater if you are deploying on premises or on the cloud. Consider that once that host comes back (could be hours later) it could be already out of date, and will replicate from the current leader, possibly dropping parts or all it's current index. Tomás > > > Tomás > > > > On Tue, Jul 5, 2016 at 11:39 AM, Lorenzo Fundaró < > > lorenzo.fund...@dawandamail.com> wrote: > > > > > @Tomas and @Steven > > > > > > I am a bit skeptical about this two statements: > > > > > > If a node just disappears you should be fine in terms of data > > > > availability, since Solr in "SolrCloud" replicates the data as it > comes > > > it > > > > (before sending the http response) > > > > > > > > > and > > > > > > > > > > > You shouldn't "need" to move the storage as SolrCloud will replicate > > all > > > > data to the new node and anything in the transaction log will already > > be > > > > distributed through the rest of the machines.. > > > > > > > > > because according to the official documentation here > > > < > > > > > > https://cwiki.apache.org/confluence/display/solr/Read+and+Write+Side+Fault+Tolerance > > > >: > > > (Write side fault tolerant -> recovery) > > > > > > If a leader goes down, it may have sent requests to some replicas and > not > > > > others. So when a new potential leader is identified, it runs a synch > > > > process against the other replicas. If this is successful, everything > > > > should be consistent, the leader registers as active, and normal > > actions > > > > proceed > > > > > > > > > I think there is a possibility that an update is not sent by the leader > > but > > > is kept in the local disk and after it comes up again it can sync the > > > non-sent data. > > > > > > Furthermore: > > > > > > Achieved Replication Factor > > > > When using a replication factor greater than one, an update request > may > > > > succeed on the shard leader but fail on one or more of the replicas. > > For > > > > instance, consider a collection with one shard and replication factor > > of > > > > three. In this case, you have a shard leader and two additional > > replicas. > > > > If an update request succeeds on the leader but fails on both > replicas, > > > for > > > > whatever reason, the update request is still considered successful > from > > > the > > > > perspective of the client. The replicas that missed the update will > > sync > > > > with the leader when they recover. > > > > > > > > > They have implemented this parameter called *min_rf* that you can use > > > (client-side) to make sure that your update was replicated to at least > > one > > > replica (e.g.: min_rf > 1). > > > > > > This is why my concern about moving storage around, because then I know > > > when the shard leader comes back, solrcloud will run sync process for > > those > > > documents that couldn't be sent to the replicas. > > > > > > Am I missing something or misunderstood the documentation ? > > > > > > Cheers ! > > > > > > > > > > > > > > > > > > > > > > > > On 5 July 2016 at 19:49, Davis, Daniel (NIH/NLM) [C] < > > daniel.da...@nih.gov > > > > > > > wrote: > > > > > > > Lorenzo, this probably comes late, but my systems guys just don't > want > > to > > > > give me real disk. Although RAID-5 or LVM on-top of JBOD may be > > better > > > > than Amazon EBS, Amazon EBS is still much closer to real disk in > terms > > of > > > > IOPS and latency than NFS ;) I even ran a mini test (not an > official > > > > benchmark), and found the response time for random reads to be > better. > > > > > > > > If you are a young/smallish company, this may be all in the cloud, > but > > if > > > > you are in a large organization like mine, you may also need to allow > > for > > > > other architectures, such as a "virtual" Netapp in the cloud that > > > > communicates with a physical Netapp on-premises, and the > > > throughput/latency > > > > of that. The most important thing is to actually measure the > numbers > > > you > > > > are getting, both for search and for simply raw I/O, or to get your > > > > systems/storage guys to measure those numbers. If you get your > > > > systems/storage guys to just measure storage - you will want to care > > > about > > > > three things for indexing primarily: > > > > > > > > Sequential Write Throughput > > > > Random Read Throughput > > > > Random Read Response Time/Latency > > > > > > > > Hope this helps, > > > > > > > > Dan Davis, Systems/Applications Architect (Contractor), > > > > Office of Computer and Communications Systems, > > > > National Library of Medicine, NIH > > > > > > > > > > > > > > > > -----Original Message----- > > > > From: Lorenzo Fundaró [mailto:lorenzo.fund...@dawandamail.com] > > > > Sent: Tuesday, July 05, 2016 3:20 AM > > > > To: solr-user@lucene.apache.org > > > > Subject: Re: deploy solr on cloud providers > > > > > > > > Hi Shawn. Actually what im trying to find out is whether this is the > > best > > > > approach for deploying solr in the cloud. I believe solrcloud solves > a > > > lot > > > > of problems in terms of High Availability but when it comes to > storage > > > > there seems to be a limitation that can be workaround of course but > > it's > > > a > > > > bit cumbersome and i was wondering if there is a better option for > this > > > or > > > > if im missing something with the way I'm doing it. I wonder if there > > are > > > > some proved experience about how to solve the storage problem when > > > > deploying in the cloud. Any advise or point to some enlightening > > > > documentation will be appreciated. Thanks. > > > > On Jul 4, 2016 18:27, "Shawn Heisey" <apa...@elyograg.org> wrote: > > > > > > > > > On 7/4/2016 10:18 AM, Lorenzo Fundaró wrote: > > > > > > when deploying solr (in solrcloud mode) in the cloud one has to > > take > > > > > > care of storage, and as far as I understand it can be a problem > > > > > > because the storage should go wherever the node is created. If we > > > > > > have for example, a node on EC2 with its own persistent disk, > this > > > > > > node happens to be the leader and at some point crashes but > > couldn't > > > > > > make the replication of the data that has in the transaction log, > > > > > > how do we do in that case ? Ideally the new node must use the > > > > > > leftover data that the death node left, but this is a bit > > cumbersome > > > > > > in my opinion. What are the best practices for this ? > > > > > > > > > > I can't make any sense of this. What is the *exact* problem you > need > > > > > to solve? The details can be very important. > > > > > > > > > > We might be dealing with this: > > > > > > > > > > http://people.apache.org/~hossman/#xyproblem > > > > > > > > > > Thanks, > > > > > Shawn > > > > > > > > > > > > > > > > > > > > > > > > > > -- > > > > > > -- > > > Lorenzo Fundaro > > > Backend Engineer > > > E-Mail: lorenzo.fund...@dawandamail.com > > > > > > Fax + 49 - (0)30 - 25 76 08 52 > > > Tel + 49 - (0)179 - 51 10 982 > > > > > > DaWanda GmbH > > > Windscheidstraße 18 > > > 10627 Berlin > > > > > > Geschäftsführer: Claudia Helming und Niels Nüssler > > > AG Charlottenburg HRB 104695 B http://www.dawanda.com > > > > > > > > > -- > > -- > Lorenzo Fundaro > Backend Engineer > E-Mail: lorenzo.fund...@dawandamail.com > > Fax + 49 - (0)30 - 25 76 08 52 > Tel + 49 - (0)179 - 51 10 982 > > DaWanda GmbH > Windscheidstraße 18 > 10627 Berlin > > Geschäftsführer: Claudia Helming und Niels Nüssler > AG Charlottenburg HRB 104695 B http://www.dawanda.com >
