[HDFS-13040] Kerberized inotify client fails despite kinit properly

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 2.6.0
• Fix Version/s: 3.1.0, 2.10.0, 3.0.3
• Component/s: namenode

Environment

Kerberized, HA cluster, iNotify client, CDH5.10.2

• Hadoop Flags: Reviewed
• Assignee: Xiao Chen
• Reporter: Wei-Chiu Chuang
• Created: 19/Jan/18 19:47
• Updated: 02/Oct/19 17:16
• Resolved: 05/Mar/18 17:48

Description

This issue is similar to HDFS-10799.

HDFS-10799 turned out to be a client side issue where client is responsible for renewing kerberos ticket actively.

However we found in a slightly setup even if client has valid Kerberos credentials, inotify still fails.

Suppose client uses principal [e-mail],
namenode 1 uses server principal hdfs/[e-mail]
namenode 2 uses server principal hdfs/[e-mail]

After Namenodes starts for longer than kerberos ticket lifetime, the client fails with the following error:

18/01/19 11:23:02 WARN security.UserGroupInformation: PriviledgedActionException as:[e-mail] (auth:KERBEROS) cause:org.apache.hadoop.ipc.RemoteException(java.io.IOException): We encountered an error reading https://nn2.example.com:8481/getJournal?jid=ns1&segmentTxId=8662&storageInfo=-60%3A353531113%3A0%3Acluster3, https://nn1.example.com:8481/getJournal?jid=ns1&segmentTxId=8662&storageInfo=-60%3A353531113%3A0%3Acluster3.  During automatic edit log failover, we noticed that all of the remaining edit log streams are shorter than the current one!  The best remaining edit log ends at transaction 8683, but we thought we could read up to transaction 8684.  If you continue, metadata will be lost forever!
        at org.apache.hadoop.hdfs.server.namenode.RedundantEditLogInputStream.nextOp(RedundantEditLogInputStream.java:213)
        at org.apache.hadoop.hdfs.server.namenode.EditLogInputStream.readOp(EditLogInputStream.java:85)
        at org.apache.hadoop.hdfs.server.namenode.NameNodeRpcServer.readOp(NameNodeRpcServer.java:1701)
        at org.apache.hadoop.hdfs.server.namenode.NameNodeRpcServer.getEditsFromTxid(NameNodeRpcServer.java:1763)
        at org.apache.hadoop.hdfs.server.namenode.AuthorizationProviderProxyClientProtocol.getEditsFromTxid(AuthorizationProviderProxyClientProtocol.java:1011)
        at org.apache.hadoop.hdfs.protocolPB.ClientNamenodeProtocolServerSideTranslatorPB.getEditsFromTxid(ClientNamenodeProtocolServerSideTranslatorPB.java:1490)
        at org.apache.hadoop.hdfs.protocol.proto.ClientNamenodeProtocolProtos$ClientNamenodeProtocol$2.callBlockingMethod(ClientNamenodeProtocolProtos.java)
        at org.apache.hadoop.ipc.ProtobufRpcEngine$Server$ProtoBufRpcInvoker.call(ProtobufRpcEngine.java:617)
        at org.apache.hadoop.ipc.RPC$Server.call(RPC.java:1073)
        at org.apache.hadoop.ipc.Server$Handler$1.run(Server.java:2216)
        at org.apache.hadoop.ipc.Server$Handler$1.run(Server.java:2212)
        at java.security.AccessController.doPrivileged(Native Method)
        at javax.security.auth.Subject.doAs(Subject.java:415)
        at org.apache.hadoop.security.UserGroupInformation.doAs(UserGroupInformation.java:1920)
        at org.apache.hadoop.ipc.Server$Handler.run(Server.java:2210)

Typically if NameNode has an expired Kerberos ticket, the error handling for the typical edit log tailing would let NameNode to relogin with its own Kerberos principal. However, when inotify uses the same code path to retrieve edits, since the current user is the inotify client's principal, unless client uses the same principal as the NameNode, NameNode can't do it on behalf of the client.

Therefore, a more appropriate approach is to use proxy user so that NameNode can retrieving edits on behalf of the client.

I will attach a patch to fix it. This patch has been verified to work for a CDH5.10.2 cluster, however it seems impossible to craft a unit test for this fix because the way Hadoop UGI handles Kerberos credentials (I can't have a single process that logins as two Kerberos principals simultaneously and let them establish connection)

A possible workaround is for the inotify client to use the active NameNode's server principal. However, that's not going to work when there's a namenode failover, because then the client's principal will not be consistent with the active NN's one, and then fails to authenticate.

Credit: this bug was confirmed and reproduced by pifta and r1pp3rj4ck

Issue Links

• is related to: HADOOP-14699 Impersonation errors with UGI after second principal relogin
• relates to: HDFS-10799 NameNode should use loginUser(hdfs) to serve iNotify requests