Talking about the HA of TajoMaster. Keeping consistence among primary master and slave masters will be a big challenge. Have we ever thought about the PAXOS protocol? It's designed to keep consistence in distributed environment.
Thanks, Daniel On Wed, Apr 16, 2014 at 7:56 PM, Hyunsik Choi <[email protected]> wrote: > Hi Alvin, > > First of all, thank you Alvin for your contribution. Your proposal looks > nice and reasonable for me. > > BTW, as other guys mentioned, TAJO-704 and TAJO-611 seem to be somewhat > overlapped to each other. We need to arrange the tasks to avoid duplicated > works. > > In my opinion, TajoMaster HA feature involves three sub features: > 1) Leader election of multiple TajoMasters - One of multiple TajoMasters > always is the leader TajoMaster. > 2) Service discovery of TajoClient side - TajoClient API call should be > resilient even though the original TajoMaster is not available. > 3) Cluster resource management and Catalog information that TajoMaster > keeps in main-memory. - the information should not be lost. > > I think that (1) and (2) are duplicated to TAJO-611 for service discovery. > So, it would be nice if TAJO-704 should only focus on (3). It's because > TAJO-611 already started few weeks ago and TAJO-704 may be the relatively > earlier stage. *Instead, you can continue the work with Xuhui and Min.* > Someone can divide the service discovery issue into more subtasks. > > In addition, I'd like to more discuss (3). Currently, a running TajoMaster > keeps two information: cluster resource information of all workers and > catalog information. In order to guarantee the HA of the data, TajoMaster > should either persistently materialize them or consistently synchronize > them across multiple TajoMasters. BTW, we will replace the resource > management feature of TajoMaster into a decentralized manner in new > scheduler issue. As a result, I think that TajoMaster HA needs to focus on > only the high availability of catalog information. The HA of catalog can be > easily achieved by database replication or we can make our own module for > it. In my view, I prefer the former. > > Hi Xuhui and Min, > > Could you share the brief progress of service discovery issue? If so, we > can easily figure out how we start the service discovery together. > > Warm regards, > Hyunsik > > > > On Wed, Apr 16, 2014 at 3:36 PM, Min Zhou <[email protected]> wrote: > > > Actually, we are not only thinking about the HA, but also service > discovery > > when the future tajo scheduler would rely on. Tajo scheduler can get all > > the active workers from that service. > > > > > > Regards, > > Min > > > > > > On Tue, Apr 15, 2014 at 10:05 PM, Xuhui Liu <[email protected]> wrote: > > > > > Hi Alvin, > > > > > > TAJO-611 will introduce Curator as a service discovery service to Tajo > > and > > > Curator is based on ZK. Maybe we can work together. > > > > > > Thanks, > > > Xuhui > > > > > > > > > On Wed, Apr 16, 2014 at 12:17 PM, Min Zhou <[email protected]> > wrote: > > > > > > > HI Alvin, > > > > > > > > I think this jira has somewhat overlap with TAJO-611, can you have > > some > > > > cooperation? > > > > > > > > Thanks, > > > > Min > > > > > > > > > > > > On Tue, Apr 15, 2014 at 7:22 PM, Henry Saputra < > > [email protected] > > > > >wrote: > > > > > > > > > Jaehwa, I think we should think about pluggable mechanism that > would > > > > > allow some kind distributed system like ZK to be used if wanted. > > > > > > > > > > - Henry > > > > > > > > > > On Tue, Apr 15, 2014 at 7:15 PM, Jaehwa Jung <[email protected]> > > > > wrote: > > > > > > Hi, Alvin > > > > > > > > > > > > I'm sorry for late response, and thank you very much for your > > > > > contribution. > > > > > > I agree with your opinion for zookeeper. But, zookeeper requires > an > > > > > > additional dependency that someone does not want. > > > > > > > > > > > > I'd like to suggest adding an abstraction layer for handling > > > TajoMaster > > > > > HA. > > > > > > When I had created TAJO-740, I wished that TajoMaster HA would > > have a > > > > > > generic interface and a basic implementation using HDFS. Next, > your > > > > > > proposed zookeeper implementation will be added there. It will > > allow > > > > > users > > > > > > to choice their desired implementation according to their > > > environments. > > > > > > > > > > > > In addition, I'd like to propose that TajoMaster embeds the HA > > > module, > > > > > and > > > > > > it would be great if HA works well by launching a backup > > TajoMaster. > > > > > > Deploying additional process besides TajoMaster and TajoWorker > > > > processes > > > > > > may give more burden to users. > > > > > > > > > > > > *Cheers* > > > > > > *Jaehwa* > > > > > > > > > > > > > > > > > > 2014-04-13 14:36 GMT+09:00 Jihoon Son <[email protected]>: > > > > > > > > > > > >> Hi Alvin. > > > > > >> Thanks for your suggestion. > > > > > >> > > > > > >> In overall, your suggestion looks very reasonable to me! > > > > > >> I'll check the POC. > > > > > >> > > > > > >> Many thanks, > > > > > >> Jihoon > > > > > >> Hi All , > > > > > >> After doing lot of research in my opinion we should > > > > utilize > > > > > >> zookeeper for Tajo Master HA.I have created a small POC and > shared > > > it > > > > > on my > > > > > >> Github repository ( [email protected]: > > alvinhenrick/zooKeeper-poc.git). > > > > > >> > > > > > >> Just to make things little bit easier and > > maintainable I > > > > am > > > > > >> utilizing Apache Curator the Fluent Zookeeper Client API > > developed > > > at > > > > > >> Netflix and is now part of an apache open source project. > > > > > >> > > > > > >> I have attached the diagram to convey my message to > > the > > > > team > > > > > >> members.Will upload it to JIRA once everyone agree with the > > proposed > > > > > >> solution. > > > > > >> > > > > > >> Here is the flow going to look like. > > > > > >> > > > > > >> TajoMasterZkController ==> > > > > > >> > > > > > >> > > > > > >> 1. This component will start and connect to zookeeper quorum > > and > > > > > fight > > > > > >> ( :) ) to obtain the latch / lock to become the master . > > > > > >> 2. Once the lock is obtained the Apache Curator API will > > > invoke > > > > > >> takeLeadership () method at this time will start the > > > TajoMaster. > > > > > >> 3. As long as the TajoMaster is running the Controller > will > > > keep > > > > > the > > > > > >> lock and update the meta data on zookeeper server with the > > > > > >> HOSTNAME and RPC > > > > > >> PORT. > > > > > >> 4. The other participant will keep waiting for the latch/ > > lock > > > > to > > > > > be > > > > > >> released by zookeeper to obtain the leadership. > > > > > >> 5. The advantage is we can have as many Tajo Master's as > we > > > > wan't > > > > > but > > > > > >> only one can be the leader and will consume the resources > > only > > > > > after > > > > > >> obtaining the latch/lock. > > > > > >> > > > > > >> > > > > > >> TajoWorkerZkController ==> > > > > > >> > > > > > >> 1. This component will start and connect to zookeeper (will > > > create > > > > > >> EPHEMERAL ZNODE) and wait for the events from zookeeper. > > > > > >> 2. The first listener will listener for successful > > > registration. > > > > > >> 3. The second listener on master node will listen for any > > > > > changes to > > > > > >> the master node received from zookeeper server. > > > > > >> 4. If the failover occurs the data on the master ZNODE > will > > > be > > > > > >> changed and the new HOSTNAME and RPC PORT can be obtained > > and > > > > the > > > > > >> TajoWorker can establish the new RPC connection with the > > > > > TajoMaster. > > > > > >> > > > > > >> To demonstrate I have created the small Readme.txt > file > > > > > >> on Github on how to run the example. Please read the log > > statements > > > on > > > > > the > > > > > >> console. > > > > > >> > > > > > >> Similar to TajoWorkerZkController we can also > > > > > >> implement TajoClientZkController. > > > > > >> > > > > > >> Any help or advice is appreciated. > > > > > >> > > > > > >> Thanks! > > > > > >> Warm Regards, > > > > > >> Alvin. > > > > > >> > > > > > > > > > > > > > > > > > > > > > -- > > > > My research interests are distributed systems, parallel computing and > > > > bytecode based virtual machine. > > > > > > > > My profile: > > > > http://www.linkedin.com/in/coderplay > > > > My blog: > > > > http://coderplay.javaeye.com > > > > > > > > > > > > > > > -- > > My research interests are distributed systems, parallel computing and > > bytecode based virtual machine. > > > > My profile: > > http://www.linkedin.com/in/coderplay > > My blog: > > http://coderplay.javaeye.com > > >
