[
https://issues.apache.org/jira/browse/HDFS-7097?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14140717#comment-14140717
]
Kihwal Lee commented on HDFS-7097:
----------------------------------
On a standby name node, the persistent state (i.e. things saved to fsimage) is
updated only by replaying edits once it is started.
- During checkpointing do not allow any state changes that will be persisted.
(I.e. edit log replaying)
- During checkpointing do not allow any RPC calls that may affect checkpointing
itself. (saveNamespace, restoreFailedStorage, etc.)
A new {{ReentrantLock}},{{nsLock}}, is introduced to coordinate checkpointing
and other activities. Any thing that requires both {{nsLock}} and {{fsLock}},
{{nsLock}} is to be locked first. Otherwise it can block other RPC calls that
do not require {{nsLock}}, not to mention deadlock. These locks are all locked
interruptibly, so that the threads can be stopped during HA transition.
Also improved {{FSImageFormat.Saver}} on cancellation checking, by making the
check interval counter a class-level variable and following the PB format's
threshold variable name. This only matters if the legacy oiv image saving is
enabled.
> Allow block reports to be processed during checkpointing on standby name node
> -----------------------------------------------------------------------------
>
> Key: HDFS-7097
> URL: https://issues.apache.org/jira/browse/HDFS-7097
> Project: Hadoop HDFS
> Issue Type: Bug
> Reporter: Kihwal Lee
> Assignee: Kihwal Lee
> Priority: Critical
> Attachments: HDFS-7097.patch
>
>
> On a reasonably busy HDFS cluster, there are stream of creates, causing data
> nodes to generate incremental block reports. When a standby name node is
> checkpointing, RPC handler threads trying to process a full or incremental
> block report is blocked on the name system's {{fsLock}}, because the
> checkpointer acquires the read lock on it. This can create a serious problem
> if the size of name space is big and checkpointing takes a long time.
> All available RPC handlers can be tied up very quickly. If you have 100
> handlers, it only takes 34 file creates. If a separate service RPC port is
> not used, HA transition will have to wait in the call queue for minutes. Even
> if a separate service RPC port is configured, hearbeats from datanodes will
> be blocked. A standby NN with a big name space can lose all data nodes after
> checkpointing. The rpc calls will also be retransmitted by data nodes many
> times, filling up the call queue and potentially causing listen queue
> overflow.
> Since block reports are not modifying any state that is being saved to
> fsimage, I propose letting them through during checkpointing.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
