[jira] [Commented] (HBASE-9426) Make custom distributed barrier procedure pluggable

["Jonathan Hsieh (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HBASE-9426?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13759085#comment-13759085 ] 

Jonathan Hsieh commented on HBASE-9426:
---------------------------------------

Before we go off and do this -- let's talk about the use cases you propose.  
* distributed log roll
* distributed table flush

What are the current mechanism used to achieve these ends, are they broken, slow, or are they sufficient?

Is the idea here to have an hbase extension mechnism for 3rd party code a la coprocessors?

At some point I'd like to change the procedure to be region-centric or possibly table-centric, as  opposed to RS-centric to make snapshotting more robust, but exposing this would make it difficult to do this evolution.  I'd suggest we have a few use cases in the core before we open this up to external plugins (a la coprocessors).
                
> Make custom distributed barrier procedure pluggable 
> ----------------------------------------------------
>
>                 Key: HBASE-9426
>                 URL: https://issues.apache.org/jira/browse/HBASE-9426
>             Project: HBase
>          Issue Type: Improvement
>    Affects Versions: 0.95.2, 0.94.11
>            Reporter: Richard Ding
>            Assignee: Richard Ding
>
> Currently if one wants to implement a custom distributed barrier procedure (e.g., distributed log roll or distributed table flush), the HBase core code needs to be modified in order for the procedure to work.
> Looking into the snapshot code (especially on region server side), most of the code to enable the procedure are generic life-cycle management (i.e., init, start, stop). We can make this part pluggable.
> Here is the proposal. Following the coprocessor example, we define two properties:
> {code}
> hbase.procedure.regionserver.classes
> hbase.procedure.master.classes
> {code}
> The values for both are comma delimited list of classes. On region server side, the classes implements the following interface:
> {code}
> public interface RegionServerProcedureManager {
>   public void initialize(RegionServerServices rss) throws KeeperException;
>   public void start();
>   public void stop(boolean force) throws IOException;
>   public String getProcedureName();
> }
> {code}
> While on Master side, the classes implement the interface:
> {code}
> public interface MasterProcedureManager {
>   public void initialize(MasterServices master) throws KeeperException, IOException, UnsupportedOperationException;
>   public void stop(String why);
>   public String getProcedureName();
>   public void execProcedure(ProcedureDescription desc) throws IOException;
>   IOException;
> }
> {code}
> Where the ProcedureDescription is defined as
> {code}
> message ProcedureDescription {
>   required string name = 1;
>   required string instance = 2;
>   optional int64 creationTime = 3 [default = 0];
>   message Property {
>     required string tag = 1;
>     optional string value = 2;
>   }
>   repeated Property props = 4;
> }
> {code}
> A generic API can be defined on HMaster to trigger a procedure:
> {code}
> public boolean execProcedure(ProcedureDescription desc) throws IOException;
> {code}
> _SnapshotManager_ and _RegionServerSnapshotManager_ are special examples of _MasterProcedureManager_ and _RegionServerProcedureManager_. They will be automatically included (users don't need to specify them in the conf file).

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira