[
https://issues.apache.org/jira/browse/HDFS-5846?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Nikola Vujic updated HDFS-5846:
-------------------------------
Status: Patch Available (was: Open)
> Assigning DEFAULT_RACK in resolveNetworkLocation method can break data
> resiliency
> ---------------------------------------------------------------------------------
>
> Key: HDFS-5846
> URL: https://issues.apache.org/jira/browse/HDFS-5846
> Project: Hadoop HDFS
> Issue Type: Bug
> Reporter: Nikola Vujic
> Assignee: Nikola Vujic
> Attachments: hdfs-5846.patch
>
>
> Medhod CachedDNSToSwitchMapping::resolve() can return NULL which requires
> careful handling. Null can be returned in two cases:
> • An error occurred with topology script execution (script crashes).
> • Script returns wrong number of values (other than expected)
> Critical handling is in the DN registration code. DN registration code is
> responsible for assigning proper topology paths to all registered datanodes.
> Existing code handles this NULL pointer on the following way
> ({{resolveNetworkLocation}} method):
> {code}
> / /resolve its network location
> List<String> rName = dnsToSwitchMapping.resolve(names);
> String networkLocation;
> if (rName == null) {
> LOG.error("The resolve call returned null! Using " +
> NetworkTopology.DEFAULT_RACK + " for host " + names);
> networkLocation = NetworkTopology.DEFAULT_RACK;
> } else {
> networkLocation = rName.get(0);
> }
> return networkLocation;
> {code}
> The line of code that is assigning default rack:
> {code} networkLocation = NetworkTopology.DEFAULT_RACK; {code}
> can cause a serious problem. This means if somehow we got NULL, then the
> default rack will be assigned as a DN's network location and DN's
> registration will finish successfully. Under this circumstances, we will be
> able to load data into cluster which is working with a wrong topology. Wrong
> topology means that fault domains are not honored.
> For the end user, it means that two data replicas can end up in the same
> fault domain and a single failure can cause loss of two, or more, replicas.
> Cluster would be in the inconsistent state but it would not be aware of that
> and the whole thing would work as if everything was fine. We can notice that
> something wrong happened almost only by looking in the log for the error:
> {code}
> LOG.error("The resolve call returned null! Using " +
> NetworkTopology.DEFAULT_RACK + " for host " + names);
> {code}
>
--
This message was sent by Atlassian JIRA
(v6.1.5#6160)
