Hi Sumit, Please see my comments inline.
On Sat, Sep 3, 2016 at 10:33 AM, sumit arrawatia <sumit.arrawa...@gmail.com> wrote: > Hi Dong, > > Please see my comments inline. > > Sumit > > On Sat, Sep 3, 2016 at 9:26 AM, Dong Lin <lindon...@gmail.com> wrote: > > > Hey Sumit, > > > > Thanks Sumit for your explanation. It seems that the concern is that we > can > > not easily change cluster.id if we read this from config. Maybe the KIP > > should also specify the requirement for users to change the cluster.id. > > > > But it seems to me that it is equally straightforward to change > cluster.id > > in both approaches. Do you think the following approach would work: > > > > *How does cluster.id <http://cluster.id> from config work:* > > > > When Kafka starts, it reads cluster.id from config. And then it reads > > cluster.id from zookeeper. > > - if the cluster.id is not specified in zookeeper, create the znode. > > - if the cluster.id is specified in zookeeper > > - if the cluster.id in znode is the same as the that in config, > proceed > > - Otherwise, broker startup fails and it reports error. Note that we > > don't make this change after the startup. > > > > This is how the current approach of generating ids would work so I agree > with you that is how setting the cluster id from config would work too. :) > > > > *How can we change cluster.id <http://cluster.id>:* > > > > - Update kafka broker config to use the new cluster.id > > - Either delete that znode from zookeeper, or update kafka broker config > to > > use a new zookeeper which doesn't have that znode > > - Do a rolling bounce > > > > With the current approach described in the KIP, if you want to change the > > cluster.id, you need to either delete the znode, or change the znode > > content, before doing a rolling bounce. I don't think the approach above > is > > more difficult than this. Any idea? > > > > I agree that this approach will work but only if we don't store cluster.id > on in meta.properties on the disk. But I think you will like the proposed > approach better if I provide some more context. > > I am not sure why you mention "... only if we don't store cluster.id in meta.properties", since neither the KIP nor my suggestion asks to store cluster.id in the meta.properties. Sorry if the confusion comes from one of my earlier where I said we can read cluster.id from meta.properties. In my proposal it should read from the same broker config file where we config zookeeper url. Anyway, since my proposed approach doesn't require the broker to store cluster.id in meta.properties, can I say we agree that user can easily change cluster.id with this approach? > I understand that your concern is that cluster ids should be human readable > and it is therefore better to let the user set it and modify it. I agree > that we should have human readable names as it makes it easier for users to > identify the cluster in their monitoring dashboard. But we also want to > allow users to change this name easily. > Yes, we all agree on this. > > At the same time, we want the cluster identity to be unique and fairly > constant for uses cases like auditing. If we allow the users to set it, we > place the burden on the users to maintain uniqueness. > I disagree that it is a burden for user to configure the unique cluster.id for each cluster. There is something that users need to configure correctly, such as zookeeper url. I don't think it is overwhelm for user to configure cluster.id correct. What we can do is to let Kafka report error if it is not configured correctly, which is covered in the approach I suggested. > So, the approach we propose is to generate a immutable UUID for the cluster > and and allow the users to assign a human readable name to the cluster > using resource tags. Tags also allow you to add structured metadata in > addition to the human readable name. For eg. if you want to identify your > cluster based on region, datacenter, env, etc. ,you can add these as tags. > Otherwise you will need to encode this in your cluster.id property. > > The current KIP lays the foundation for this approach. We plan to implement > the first part (UUID) in a manner that makes it easy to add the second part > (human readable names using tags) in a future KIP. > Are you suggesting that it is better to use randomly generated cluster.id + human-readable tag than using human-readable cluster.id? If so, can you explain the reason? > > Hopefully this helps to clarify your doubts and concerns. > > > > Cheers, > > Dong > > > > > > > > On Sat, Sep 3, 2016 at 12:46 AM, sumit arrawatia < > > sumit.arrawa...@gmail.com> > > wrote: > > > > > Hi Dong, > > > > > > If you set the cluster.id in the config, the problem is how you > > > change/update the cluster.id . > > > > > > > > > You will need to change the all the config files and make sure every > one > > of > > > them is correct as well as update the ZK metadata. This will require a > > > reboot/downtime of the entire cluster, whereas generating ids (along > with > > > the yet-to-be-published resource tags KIP ) makes it easy for the > admins > > to > > > update the human readable name without reboots/ clusterwide changes to > > > config. > > > > > > > > > Also, when we implement the change to write cluster.id in > > meta.properties, > > > the process of updating the cluster.id becomes even more complicated. > > > > > > > > > 1. We will have to ensure that the entire process is transactional. It > > will > > > be painful from an operational standpoint as it would require as much > > > operational downtime and care as a version upgrade because it will > modify > > > the on-disk data. The recovery/rollback scenario would be very > difficult > > > and would probably need manual changes to meta.properties. > > > > > > > > > 2. Given that we would want to generate an error if the cluster.id in > > > meta.properties doesn't match cluster.id in ZK, we will have to setup > > > complicated logic in ZK to ensure we forgo the check when changing the > > > cluster.id. I am not even sure how to do it properly for a rolling > > upgrade > > > without downtime. > > > > > > > > > All these points are based on my experience running Elasticsearch in > > > production. ElasticSearch specifies cluster.name statically in the > > > properties as well as includes it in the data directory name and > changing > > > it is a nightmare. > > > > > > > > > You would think that naming changes should be rare but in my experience > > > they are not. Sometimes typos creep in, sometimes you have to change > the > > > name to consolidate clusters or divide them and sometimes the infra > team > > > decides to change the deployment metadata. > > > > > > > > > This is why I think AWS approach of (assigning immutable ids to > > resources + > > > human readable names in tags) works very well operationally. > > > > > > > > > Hope this helps ! > > > > > > > > > Cheers, > > > > > > Sumit > > > > > > On Fri, Sep 2, 2016 at 11:44 PM, Dong Lin <lindon...@gmail.com> wrote: > > > > > > > Hey Ismael, > > > > > > > > Thanks for your reply. Please see my comment inline. > > > > > > > > On Fri, Sep 2, 2016 at 8:28 PM, Ismael Juma <ism...@juma.me.uk> > wrote: > > > > > > > > > Hi Dong, > > > > > > > > > > Thanks for your feedback. Comments inline. > > > > > > > > > > On Thu, Sep 1, 2016 at 7:51 PM, Dong Lin <lindon...@gmail.com> > > wrote: > > > > > > > > > > > > I share the view with Harsha and would like to understand how the > > > > current > > > > > > approach of randomly generating cluster.id compares with the > > > approach > > > > of > > > > > > manually specifying it in meta.properties. > > > > > > > > > > > > > > > > Harsha's suggestion in the thread was to store the generated id in > > > > > meta.properties, not to manually specify it via meta.properties. > > > > > > > > > > > > > > > > > I think one big advantage of defining it manually in zookeeper is > > > that > > > > we > > > > > > can easily tell which cluster it is by simply looking at the > sensor > > > > name, > > > > > > which makes it more useful to the auditing or monitoring use-case > > > that > > > > > this > > > > > > KIP intends to address. > > > > > > > > > > > > > > > If you really want to customise the name, it is possible with the > > > current > > > > > proposal: save the appropriate znode in ZooKeeper before a broker > > > > > auto-generates it. We don't encourage that because once you have a > > > > > meaningful name, there's a good chance that you may want to change > it > > > in > > > > > the future. And things break down at that point. That's why we > prefer > > > > > having a generated, unique and immutable id complemented by a > > > changeable > > > > > human readable name. As described in the KIP, we think the latter > can > > > be > > > > > achieved more generally via resource tags (which will be a separate > > > KIP). > > > > > > > > > > Can you elaborate what will break down if we need to change the > name? > > > > > > > > Even if we can not change name because something will breakdown in > that > > > > case, it seems that it is still better to read id from config than > > using > > > a > > > > randomly generated ID. In my suggested solution user can simply > choose > > > not > > > > to change the name and make sure there is unique id per cluster. In > > your > > > > proposal you need to store the old cluster.id and manually restore > it > > in > > > > zookeeper in some scenarios. What do you think? > > > > > > > > > > > > > > On the other hand, if you can only tell whether two > > > > > > sensors are measuring the same cluster or not. Also note that > even > > > this > > > > > > goal is not easily guaranteed, because you need an external > > mechanism > > > > to > > > > > > manually re-generate znode with the old cluster.id if znode is > > > deleted > > > > > or > > > > > > if the same cluster (w.r.t purpose) is changed to use a different > > > > > > zookeeper. > > > > > > > > > > > > > > > > If we assume that znodes can be deleted at random, the cluster id > is > > > > > probably the least of one's worries. And yes, when moving to a > > > > > different ZooKeeper while wanting to retain the cluster id, you > would > > > > have > > > > > to set the znode manually. This doesn't seem too onerous compared > to > > > the > > > > > other work you will have to do for this scenario. > > > > > > > > > > Maybe this work is not much compared to other work. But we can > agree > > > that > > > > no work is better than little work, right? I am interested to see if > we > > > can > > > > avoid the work and still meet the motivation and goals of this KIP. > > > > > > > > > > > > > > I read your reply to Harsha but still I don't fully understand > your > > > > > concern > > > > > > with that approach. I think the broker can simply register > > group.id > > > in > > > > > > that > > > > > > znode if it is not specified yet, in the same way that this KIP > > > > proposes > > > > > to > > > > > > do it, right? Can you please elaborate more about your concern > with > > > > this > > > > > > approach? > > > > > > > > > > > > > > > > It's a bit difficult to answer this comment because it seems like > the > > > > > intent of your suggestion is different than Harsha's. > > > > > > > > > > I am not necessarily opposed to storing the cluster id in > > > meta.properties > > > > > (note that we have one meta.properties per log.dir), but I think > > there > > > > are > > > > > a number of things that need to be discussed and I don't think we > > need > > > to > > > > > block KIP-78 while that takes place. Delivering features > > incrementally > > > > is a > > > > > good thing in my opinion (KIP-31/32, KIP-33 and KIP-79 is a good > > recent > > > > > example). > > > > > > > > > > > > > If I understand it right, the motivation of this KIP is to allow > > cluster > > > to > > > > be uniquely identified. This is a useful feature and I am not asking > > for > > > > anything beyond this scope. It is just that reading cluster.id from > > > config > > > > seems to be a better solution in order to meet the motivation and all > > the > > > > goals described in the KIP. More specifically, using cluster.id not > > only > > > > allows user to distinguish between different clusters, it also lets > > user > > > > identify cluster. In comparison, randomly generated cluster.id > allows > > > user > > > > to distinguish cluster with a little bit more effort, and doesn't > allow > > > > user to identify a cluster by simply reading e.g. sensor name. Did I > > miss > > > > something here? > > > > > > > > > > > > > > > > > > Ismael > > > > > > > > > > P.S. For what is worth, the following version of the KIP includes > an > > > > > incomplete description (it assumes a single meta.properties, but > > there > > > > > could be many) of what the broker would have to do if we wanted to > > save > > > > to > > > > > meta.properties and potentially restore the znode from it. The > state > > > > space > > > > > becomes a lot more complex, increasing potential for bugs (we had a > > few > > > > for > > > > > generated broker ids). In contrast, the current proposal is very > > simple > > > > and > > > > > doesn't prevent us from introducing the additional functionality > > later. > > > > > > > > > > https://cwiki.apache.org/confluence/pages/viewpage. > > > > action?pageId=65868433 > > > > > > > > > > > > > IMO reading cluster.id from config should be as easy as reading > broker > > > id > > > > from config. Storing cluster.id from config in znode requires the > same > > > > amount of effort as storing randomly generated cluster.id in znode. > > > Maybe > > > > I > > > > missed something here. Can you point me to the section of the KIP > that > > > > explains why it is more difficult if we want to read cluster.id from > > > > config? > > > > > > > > > > > > > > > > -- > > > Regards, > > > Sumit > > > > > > > > > -- > Regards, > Sumit >
