[
https://issues.apache.org/jira/browse/HBASE-18846?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
stack updated HBASE-18846:
--------------------------
Attachment: hbase-site.xml
IndexerConnection.java
Example hbase-site.xml for hbase-indexer to use; shuts down all RegionServer
Services and installs an alternate Connection implementation, also attached,
which shows where you'd inject code to catch a Replication Stream.
> Accommodate the hbase-indexer/lily/SEP consumer deploy-type
> -----------------------------------------------------------
>
> Key: HBASE-18846
> URL: https://issues.apache.org/jira/browse/HBASE-18846
> Project: HBase
> Issue Type: Bug
> Reporter: stack
> Assignee: stack
> Fix For: 2.0.0-beta-1
>
> Attachments: HBASE-18846.master.001.patch,
> HBASE-18846.master.002.patch, HBASE-18846.master.003.patch,
> IndexerConnection.java, hbase-site.xml, javadoc.txt
>
>
> This is a follow-on from HBASE-10504, Define a Replication Interface. There
> we defined a new, flexible replication endpoint for others to implement but
> it did little to help the case of the lily hbase-indexer. This issue takes up
> the case of the hbase-indexer.
> The hbase-indexer poses to hbase as a 'fake' peer cluster (For why
> hbase-indexer is implemented so, the advantage to having the indexing done in
> a separate process set that can be independently scaled, can participate in
> the same security realm, etc., see discussion in HBASE-10504). The
> hbase-indexer will start up a cut-down "RegionServer" processes that are just
> an instance of hbase RpcServer hosting an AdminProtos Service. They make
> themselves 'appear' to the Replication Source by hoisting up an ephemeral
> znode 'registering' as a RegionServer. The source cluster then streams
> WALEdits to the Admin Protos method:
> {code}
> public ReplicateWALEntryResponse replicateWALEntry(final RpcController
> controller,
> final ReplicateWALEntryRequest request) throws ServiceException {
> {code}
> The hbase-indexer relies on other hbase internals like Server so it can get a
> ZooKeeperWatcher instance and know the 'name' to use for this cut-down server.
> Thoughts on how to proceed include:
>
> * Better formalize its current digestion of hbase internals; make it so
> rpcserver is allowed to be used by others, etc. This would be hard to do
> given they use basics like Server, Protobuf serdes for WAL types, and
> AdminProtos Service. Any change in this wide API breaks (again)
> hbase-indexer. We have made a 'channel' for Coprocessor Endpoints so they
> continue to work though they use 'internal' types. They can use protos in
> hbase-protocol. hbase-protocol protos are in a limbo currently where they are
> sort-of 'public'; a TODO. Perhaps the hbase-indexer could do similar relying
> on the hbase-protocol (pb2.5) content and we could do something to reveal
> rpcserver and zk for hbase-indexer safe use.
> * Start an actual RegionServer only have it register the AdminProtos Service
> only -- not ClientProtos and the Service that does Master interaction, etc.
> [I checked, this is not as easy to do as I at first thought -- St.Ack] Then
> have the hbase-indexer implement an AdminCoprocessor to override the
> replicateWALEntry method (the Admin CP implementation may need work). This
> would narrow the hbase-indexer exposure to that of the Admin Coprocessor
> Interface
> * Over in HBASE-10504, [~enis] suggested "... if we want to provide
> isolation for the replication services in hbase, we can have a simple host as
> another daemon which hosts the ReplicationEndpoint implementation. RS's will
> use a built-in RE to send the edits to this layer, and the host will delegate
> it to the RE implementation. The flow would be something like: RS --> RE
> inside RS --> Host daemon for RE --> Actual RE implementation --> third party
> system..."
>
> Other crazy notions occur including the setup of an Admin Interface
> Coprocessor Endpoint. A new ReplicationEndpoint would feed the replication
> stream to the remote cluster via the CPEP registered channel.
> But time is short. Hopefully we can figure something that will work in 2.0
> timeframe w/o too much code movement.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
