I looked into the consumer offset storage and it seems like for acl storage we should not need something as complex. Consumer offset has different throughput requirements which is why I think it made sense to move away from zookeeper. Acls on the other hand seldom change and because of the caching layer will require very low read rate. Zookeeper seems like a perfect storage solution for small metadata like this.
As far as mirror maker goes, given we are adding the storage APIs for acls (add/remove/get) it should be easy for mirror maker to use these APIs to read acls and add it back in the new cluster. Thanks Parth On 4/16/15, 6:13 PM, "Jun Rao" <j...@confluent.io> wrote: >Hi, Gwen, > >What you suggested seems reasonable. I guess we will need the <Principal, >Privilege> pair and the Resource in grant() and revoke()? > >Is the Hive authorization api the following? It's weird that it takes user >in checkPermissions(), but not in authorize(). > >http://hive.apache.org/javadocs/r0.11.0/api/org/apache/hadoop/hive/ql/secu >rity/authorization/StorageBasedAuthorizationProvider.html > >I was imagining that a default implementation could be similar to how we >store offsets in Kafka. Basically, store all acls in a special topic with >compact retention. Then, every broker will build an in-memory cache off >that topic. > >Another thing that we haven't discussed so far is how to manage ACLs >across >different mirrored Kafka clusters. Let's say you use mirror maker to >mirror >all topics from cluster A to cluster B. You probably want to have exactly >the same ACL on both A and B. It would be good if the ACL can be set up >just once. If we use the above default implementation, since the ACL topic >is mirrored too, the ACL will be propagated automatically. > >Thanks, > >Jun > > >On Thu, Apr 16, 2015 at 9:44 AM, Gwen Shapira <gshap...@cloudera.com> >wrote: > >> Hi Kafka Authorization Fans, >> >> I'm starting a new thread on a specific sub-topic of KIP-11, since >> this is a bit long :) >> >> Currently KIP-11, as I understand it, proposes: >> * Authorizers are pluggable, with Kafka providing DefaultAuthorizer. >> * Kafka tools allow adding / managing ACLs. >> * Those ACLs are stored in ZK and cached in a new TopicCache >> * Authorizers can either use the ACLs defined and stored in Kafka, or >> define and use their own. >> >> I am concerned of two possible issues with this design: >> 1. Separation of concerns - only authorizers should worry about ACLs, >> and therefore the less code for ACLs that exist in Kafka core, the >> better. >> 2. User confusion - It sounded like we can define ACLs in Kafka itself >> but authorizers can also define their own, so "kafka-topics >> --describe" may show an ACL different than the one in use. This can be >> super confusing for admins. >> >> My alternative suggestion: >> * Authorizer API will include: >> grantPrivilege(List<Principals>, List<Privilege>) >> revokePrivilege(List<Principals>, List<Privilege>), >> getPrivilegesByPrincipal(Principal, Resource) >> .... >> (The exact API can be discussed in detail, but you get the idea) >> * Kafka tools will simply invoke these APIs when topics are added / >> modified / described. >> * Each authorizer (including the default one) will be responsible for >> storing, caching and using those ACLs. >> >> This way, we keep almost all ACL code with the Authorizer, where it >> belongs and users get a nice unified interface that reflects what is >> actually getting used in the system. >> This is pretty much how Sqoop and Hive implement their authorization >>APIs. >> >> What do you think? >> >> Gwen >>
