[
https://issues.apache.org/jira/browse/HDFS-8693?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14615361#comment-14615361
]
Jian Fang commented on HDFS-8693:
---------------------------------
[~kihwal], thank you for your comments. Unfortunately, DNS alias does not
really work or comes with a very high maintenance cost in some scenarios when
the cluster is not maintained by system admin, but by software such as the case
of auto provisioning and managing Hadoop cluster in cloud. HDFS clients have
the failover proxy configured to switch to one of the two pre-configured name
node addresses. There are cases the new name node becomes active and the HDFS
clients are broken, but we more concern about the data nodes themselves for a
long running cluster. The refreshNameNodes needs to be fixed in this case.
> refreshNamenodes does not support adding a new standby to a running DN
> ----------------------------------------------------------------------
>
> Key: HDFS-8693
> URL: https://issues.apache.org/jira/browse/HDFS-8693
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: datanode, ha
> Affects Versions: 2.6.0
> Reporter: Jian Fang
> Priority: Critical
>
> I tried to run the following command on a Hadoop 2.6.0 cluster with HA
> support
> $ hdfs dfsadmin -refreshNamenodes datanode-host:port
> to refresh name nodes on data nodes after I replaced one name node with a new
> one so that I don't need to restart the data nodes. However, I got the
> following error:
> refreshNamenodes: HA does not currently support adding a new standby to a
> running DN. Please do a rolling restart of DNs to reconfigure the list of NNs.
> I checked the 2.6.0 code and the error was thrown by the following code
> snippet, which led me to this JIRA.
> void refreshNNList(ArrayList<InetSocketAddress> addrs) throws IOException {
> Set<InetSocketAddress> oldAddrs = Sets.newHashSet();
> for (BPServiceActor actor : bpServices)
> { oldAddrs.add(actor.getNNSocketAddress()); }
> Set<InetSocketAddress> newAddrs = Sets.newHashSet(addrs);
> if (!Sets.symmetricDifference(oldAddrs, newAddrs).isEmpty())
> { // Keep things simple for now -- we can implement this at a later date.
> throw new IOException( "HA does not currently support adding a new standby to
> a running DN. " + "Please do a rolling restart of DNs to reconfigure the list
> of NNs."); }
> }
> Looks like this the refreshNameNodes command is an uncompleted feature.
> Unfortunately, the new name node on a replacement is critical for auto
> provisioning a hadoop cluster with HDFS HA support. Without this support, the
> HA feature could not really be used. I also observed that the new standby
> name node on the replacement instance could stuck in safe mode because no
> data nodes check in with it. Even with a rolling restart, it may take quite
> some time to restart all data nodes if we have a big cluster, for example,
> with 4000 data nodes, let alone restarting DN is way too intrusive and it is
> not a preferable operation in production. It also increases the chance for a
> double failure because the standby name node is not really ready for a
> failover in the case that the current active name node fails.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
