Erik Krogen created HDFS-13689:
----------------------------------
Summary: NameNodeRpcServer getEditsFromTxid assumes it is run on
active NameNode
Key: HDFS-13689
URL: https://issues.apache.org/jira/browse/HDFS-13689
Project: Hadoop HDFS
Issue Type: Sub-task
Components: hdfs, namenode
Reporter: Erik Krogen
Assignee: Erik Krogen
{{NameNodeRpcServer#getEditsFromTxid}} currently decides which transactions are
able to be served, i.e. which transactions are durable, using the following
logic:
{code}
long syncTxid = log.getSyncTxId();
// If we haven't synced anything yet, we can only read finalized
// segments since we can't reliably determine which txns in in-progress
// segments have actually been committed (e.g. written to a quorum of JNs).
// If we have synced txns, we can definitely read up to syncTxid since
// syncTxid is only updated after a transaction is committed to all
// journals. (In-progress segments written by old writers are already
// discarded for us, so if we read any in-progress segments they are
// guaranteed to have been written by this NameNode.)
boolean readInProgress = syncTxid > 0;
{code}
This assumes that the NameNode serving this request is the current
writer/active NameNode, which may not be true in the ObserverNode situation.
Since {{selectInputStreams}} now has a {{onlyDurableTxns}} flag, which, if
enabled, will only return durable/committed transactions, we can instead
leverage this to provide the same functionality. We should utilize this to
avoid consistency issues when serving this request from the ObserverNode.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: hdfs-dev-unsubscr...@hadoop.apache.org
For additional commands, e-mail: hdfs-dev-h...@hadoop.apache.org
