[
https://issues.apache.org/jira/browse/HBASE-1157?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12682083#action_12682083
]
Jim Kellerman commented on HBASE-1157:
--------------------------------------
Sorry, meant to include the list of changes (which I had written down), but
forgot. Here they are:
Summary of change:
Server name is now hostname_startcode_port so it is unique among multiple
instances of a region server running at the same address
Files changed:
HServerInfo
- Cache server name once it's been computed so that subsequent gets of the
server name are a simple dereference. This required setServerAddress,
setStartCode to be synchronized as well as getServerName.
- Return a copy of the embedded HServerAddress since an HRegionInfo can get
reused and the reference returned can change out from other methods that have
stored that reference.
- Changed hashCode, compareTo to just use the serverName as that is now unique.
- new static methods to compute serverName given a HServerInfo; hostName
(name:port String) and startCode; and one that takes a HServerAddress and
startCode.
HRegionInfo
- Not actually needed for this issue, but I had at one point thought I needed
to change it, and when I discarded the change, I left the following in
- make regionName transient as it is not serialized
- fix javadoc for parseRegionName, shouldSplit
HMaster
- regionServerStartup can now throw an exception which comes out of
ServerManager.regionServerStartup
ServerManager
- remove unused import
- the key for serversToServerInfo is still a string but is the server name as
described above, and not host:port as it was previously. This allows us to
distinguish between different instances of a region server at the same address.
Zookeeper still uses a ServerExpirer object, but the name is now the server
name as described above rather than a host:port pair.
- removed Boolean from deadServers Map. It was not used and is no longer
needed. deadServers is now a set. deadServers is still used to detect a server
that has not restarted but reports in after its lease has expired (perhaps due
to a network partitioning). It cannot be given work until its existing logs
have been recovered.
- removed checkForGhostReferences. deadServers now handles servers that report
back in after lease expiration but have not restarted. Servers that have
restarted will have a new server name and can consequently be assigned work
immediately.
- regionServerStartup can now throw Leases.LeaseStillHeldException. Changed
String 's' to String 'serverName'.
- regionServerReport can now throw Leases.LeaseStillHeldException.
- processRegionServerExit, processRegionServerAllsWell, processMsgs,
processRegionOpen now take a HServerInfo rather than a string server name
- processSplitRegion no longer takes either a server name or a serverInfo
because they were unused.
- getServersToServerInfo, getServersToLoad now return a
Collections.unmodifyableMap instead of a new HashMap
- getLoadToServers removed. Unreferenced.
RegionManager
- regionsInTransition is now a Map<String, RegionState> instead of Map<byte[],
RegionState> key is server name described above
- assignRegions now just takes an HServerInfo instead of HServerInfo, String as
that was redundant and the server name can be obtained from HServerInfo
- unassignSomeRegions, assignRegionsToOneServer, assignRegionsToMultipleServers
take an HServerInfo rather than String
- regionIsInTransition, regionIsOpening, isPendingOpen, setOpen, isOfflined,
setPendingClose, setClosed take String as argument instead of byte[] due to
change in regionsInTransition map
BaseScanner
- checkAssigned now composes a server name as described above from
serverAddress and startCode. Additionally it no longer needs to compare start
codes as ServerManager.getServerInfo uses a server name as described above and
since nothing will be found if the names do not match exactly (since the
startcode is part of the name) the comparison of start codes is redundant.
Finally it calls new static function HLog.getHLogDirectoryName instead of
duplicating the code.
TableOperation
- now builds a server name as described above to pass to isBeingServed and
processScanItem
- isBeingServed, processScanItem now take a server name as described above
rather than host:port and start code
ChangeTableState
- processScanItem now takes a server name as described above rather than
host:port and start code
ModifyTableMeta
- added @SuppressWarnings("unused") for server name argument in processScanItem
TableDelete
- added @SuppressWarnings("unused") for server name argument in processScanItem
ColumnOperation
- added @SuppressWarnings("unused") for server name argument in processScanItem
ProcessRegionOpen
- now stores the HServerInfo passed to the constructor rather than the
HServerAddress and start code contained therein
- added if (LOG.isDebugEnabled()) { around LOG.debug calls
- other changes are related to extracting the HServerAddress and start code
fields from the HServerInfo
ProcessServerShutdown
- no longer needs the HServerAddress as it constructs a server name as
described above for comparision with entries scanned from meta region.
HLog
- Added new static methods getHLogDirectoryName that take either a HServerInfo
a host:port string and start code or a string which is a server name as
described above
> If we do not take start code as a part of region server recovery, we could
> inadvertantly try to reassign regions assigned to a restarted server with a
> different start code
> ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------
>
> Key: HBASE-1157
> URL: https://issues.apache.org/jira/browse/HBASE-1157
> Project: Hadoop HBase
> Issue Type: Bug
> Components: master, regionserver
> Affects Versions: 0.20.0
> Reporter: Jim Kellerman
> Assignee: Jim Kellerman
> Fix For: 0.20.0
>
>
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
