[
https://issues.apache.org/jira/browse/HBASE-1736?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12846074#action_12846074
]
stack commented on HBASE-1736:
------------------------------
This issue is not as bad as it was. We made a change so that if the RS fails
to deliver message to master, it'll retry rather than drop the message on the
ground as it used to.
There is also the notion that loss of a split, while a pain, is not the end of
the world. We can redo it if we have to. There is some comment in the code in
CompactSplitThread#split where try to reason what happens to split if crash at
various points along the split. They could do with review I'd say (smile).
On "It should add the children to the META before closing the parent, so that
children are not lost if RS crashes after closing parent", what are you
thinking? The parent would still be online when we add in the daughter regions
-- pre-close? We'd have to deal then with clients coming in and finding the
daughter regions instead of the parent in .META. (See how
HRegion#getClosestRowBefore works and then how its used in
HCM#locateRegionInMeta). We might have to add in some code to not hand out
daughter regions that don't have an info:server (They won't have an info:server
value in .META. if they have not deployed) but this will still be
unsatisfactory since most of the time it'll just end up with an offlined parent.
A related, big issue in here is currently closing the parent can take some time
during which attempts at reaching the parent region are rejected with
NotServingRegionException. Parents can take a while to close because flush of
memory content happens while the 'closing' flag is set on the region. Its as
though we need to flush before we go into the close taking on writes while we
do so and then only flush the small amount we accumulated during the flush
while under the 'closing' flag (This is still unsatisfactory because if loaded
system and flushes are slow, memory might be filled by the time flush completes
and again we'll have a slow close because we're flushing lots of memory).
There is a 'make splits faster' issue already. I'm just calling attention to
it here since it a little related.
> If RS can't talk to master, pause; more importantly, don't split (Currently
> we do and splits are lost and table is wounded)
> ---------------------------------------------------------------------------------------------------------------------------
>
> Key: HBASE-1736
> URL: https://issues.apache.org/jira/browse/HBASE-1736
> Project: Hadoop HBase
> Issue Type: Bug
> Reporter: stack
> Assignee: stack
> Priority: Critical
> Fix For: 0.21.0
>
>
> What I saw was master shutting itself down because it had lost zk lease.
> Fine. The RS though doesn't look like it can deal with this situation.
> We'll see stuff like this:
> {code}
> ...failed on connection exception: java.net.ConnectException: Connection
> refused
> at
> org.apache.hadoop.hbase.ipc.HBaseClient.wrapException(HBaseClient.java:744)
> at org.apache.hadoop.hbase.ipc.HBaseClient.call(HBaseClient.java:722)
> at org.apache.hadoop.hbase.ipc.HBaseRPC$Invoker.invoke(HBaseRPC.java:328)
> at $Proxy0.regionServerReport(Unknown Source)
> at
> org.apache.hadoop.hbase.regionserver.HRegionServer.run(HRegionServer.java:470)
> at java.lang.Thread.run(Unknown Source)
> Caused by: java.net.ConnectException: Connection refused
> at sun.nio.ch.SocketChannelImpl.checkConnect(Native Method)
> at sun.nio.ch.SocketChannelImpl.finishConnect(Unknown Source)
> at
> org.apache.hadoop.net.SocketIOWithTimeout.connect(SocketIOWithTimeout.java:206)
> at org.apache.hadoop.net.NetUtils.connect(NetUtils.java:404)
> at
> org.apache.hadoop.hbase.ipc.HBaseClient$Connection.setupIOstreams(HBaseClient.java:305)
> at
> org.apache.hadoop.hbase.ipc.HBaseClient.getConnection(HBaseClient.java:826)
> at org.apache.hadoop.hbase.ipc.HBaseClient.call(HBaseClient.java:707)
> ... 4 more
> {code}
> ... all over the regionserver as it tries to send heartbeat to master on this
> broken connection.
> On split, we close parent, add children to the catalog but then when we try
> to tell the master about the split, it fails. Means the children never get
> deployed. Meantime the parent is offline.
> This issue is about going through the regionserver and anytime it has a
> connection to master, make sure on fault that no damage is done the table and
> then that the regionserver puts a pause on splitting.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
