[
https://issues.apache.org/jira/browse/HBASE-19114?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16226965#comment-16226965
]
Appy commented on HBASE-19114:
------------------------------
bq. ... most zk related code will be in hbase-server...
That's what we are trying to move away from- everything in hbase-server. It has
60% of our codebase - and that too in a very tangled and bad state.
Since packages don't impose circular dependency violations, everything within a
single module can depend on everything else. And that has lead this nasty
situation.
(https://docs.google.com/document/d/1wZAimGcJzc0jys0-EATRi0CyGMVlcIxTvZtegPF4mfw/edit)
Those dashed edges are backward dependency and so many of them is a clear
signal that we need to design and place our code more appropriately, like in
recent times with separate hbase-backup module, hbase-replication module.
bq. We already have too many modules...
I don't know if 10 is too many, or 20, or 30. 100 definitely seems too many. 2
seems too low. Is there a metrics on which we can judge that? Because am
gunning towards straightening our dependencies graph further by splitting out
hbase-io and hbase-wal. Advantages are mentioned in that doc mentioned above.
I guess the high level thing here is, anything thats directly zookeeper
dependent and doesn't require our internals can be easily split out and kept
separately from our internals.
> Split out o.a.h.h.zookeeper from hbase-server and hbase-client
> --------------------------------------------------------------
>
> Key: HBASE-19114
> URL: https://issues.apache.org/jira/browse/HBASE-19114
> Project: HBase
> Issue Type: Sub-task
> Reporter: Appy
> Assignee: Appy
> Attachments: HBASE-19114.master.001.patch,
> HBASE-19114.master.002.patch, HBASE-19114.master.003.patch
>
>
> Changes so far:
> - Moved DrainingServerTracker and RegionServerTracker to
> hbase-server:o.a.h.h.master.
> - Move Abortable to hbase-common. Since it's IA.Private and independent of
> anything, moving it to hbase-common which is at bottom of the dependency tree
> is better.
> - Moved RecoveringRegionWatcher to hbase-server:o.a.h.h.regionserver
> - Moved SplitOrMergeTracker to oahh.master (because it depends on a PB)
> - Moving hbase-client:oahh.zookeeper.* to hbase-zookeeper module. We want to
> keep hbase-zookeeper very independent and hence at lowest levels in our
> dependency tree.
> - ZKUtil is a huge tangle since it's linked to almost everything in
> \[hbase-client/]oahh.zookeeper. And pulling it down requires some basic proto
> functions (mergeFrom, PBmagic, etc). So what i did was:
> ** Pulled down common and basic protobuf functions (which only depend on
> com.google.protobuf.\*) to hbase-common so other code depending on them can
> be pulled down if possible/wanted in future. This will help future dependency
> untangling too. These are ProtobufMagic and ProtobufHelpers.
> ** Didn't move any hbase-specific PB stuff to hbase-common. We can't pull
> things into hbase-common which add dependency between it and
> hbase-protobuf/hbase-shaded-protobuf since we very recently untangled them.
> - DEFAULT_REPLICA_ID is used in many places in ZK. Declared a new constant in
> HConstants (since it's in hbase-common) and using it in hbase-zookeeper.
> RegionInfo.DEFAULT_REPLICA_ID also takes its value from it to avoid case
> where two values can become different.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
