[HDFS-1725] Set storage directories only at FSImage construction (was Cleanup FSImage construction)

• Type: Sub-task
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: Edit log branch (HDFS-1073)
• Hadoop Flags: Reviewed
• Assignee: Ivan Kelly
• Reporter: Ivan Kelly
• Created: 04/Mar/11 12:11
• Updated: 18/May/11 18:48
• Resolved: 18/May/11 18:48

Description

HDFS-1580 proposes extending FSEditLog to allow it to use editlog streams which are not backed by StorageDirectory. Currently, to set the the directories used for edits, NNStorage#setStorageDirectory is called with a list of URIs as the second argument. NNStorage takes this list or URIs, takes all file:/// URIs and adds them to its StorageDirectory list. Then, when opened, FSEditLog will request a list of StorageDirectories from NNStorage and create a list of EditLogOutputStreams based on these.

This approach cannot work with HDFS-1580. NNStorage exists solely to deal with filesystem based storage. As such, only StorageDirectories can be retrieved from NNStorage by FSEditLog. So, FSEditLog should get the URI from some place other than NNStorage. This presents a further problem, in that, NNStorage#setStorageDirectories is the current way of setting the URIs for images and edits. This call can happen at any time, so the directories in NNStorage can change at any time. If FSEditLog is to get its URIs from elsewhere, this opens up the risk of the filesystem directories in NNStorage and filesystem URIs being out of sync.

A solution to this is to stipulate that the URIs for NNStorage are set only once, on construction. All proper uses of NNStorage#setStorageDirectories are being called just after construction of the image in any case. All other cases are using NNStorage#setStorageDirectories not to set the storage directories, but for the side effects of this call. This guide explains these other cases.