[
https://issues.apache.org/jira/browse/HBASE-14090?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14629516#comment-14629516
]
Lars Hofhansl commented on HBASE-14090:
---------------------------------------
To throw out a radical idea: We _could_ go as far as having HBase manage a
block pool directly, without the NN being involved at all. We'd lose things
like distCP, etc, but we'd be in control of all blocks directly (placement,
etc), and no need to run a pesky NN (with QJMs, etc). Let me dream for a bit :)
> Redo FS layout; let go of tables/regions/stores directory hierarchy in DFS
> --------------------------------------------------------------------------
>
> Key: HBASE-14090
> URL: https://issues.apache.org/jira/browse/HBASE-14090
> Project: HBase
> Issue Type: Sub-task
> Reporter: stack
>
> Our layout as is won't work if 1M regions; e.g. HDFS will fall over if
> directories of hundreds of thousands of files. HBASE-13991 (Humongous Tables)
> would address this specific directory problem only by adding subdirs under
> table dir but there are other issues with our current layout:
> * Our table/regions/column family 'facade' has to be maintained in two
> locations -- in master memory and in the hdfs directory layout -- and the
> farce needs to be kept synced or worse, the model management is split between
> master memory and DFS layout. 'Syncing' in HDFS has us dropping constructs
> such as 'Reference' and 'HalfHFiles' on split, 'HFileLinks' when archiving,
> and so on. This 'tie' makes it hard to make changes.
> * While HDFS has atomic rename, useful for fencing and for having files
> added atomically, if the model were solely owned by hbase, there are hbase
> primitives we could make use of -- changes in a row are atomic and
> coprocessors -- to simplify table transactions and provide more consistent
> views of our model to clients; file 'moves' could be a memory operation only
> rather than an HDFS call; sharing files between tables/snapshots and when it
> is safe to remove them would be simplified if one owner only; and so on.
> This is an umbrella blue-sky issue to discuss what a new layout would look
> like and how we might get there. I'll follow up with some sketches of what
> new layout could look like that come of some chats a few of us have been
> having. We are also under the 'delusion' that move to a new layout could be
> done as part of a rolling upgrade and that the amount of work involved is not
> gargantuan.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
