[
https://issues.apache.org/jira/browse/HBASE-17375?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15779498#comment-15779498
]
Chang chen commented on HBASE-17375:
------------------------------------
oh my, I didn't see this on [http://hbase.apache.org/book.html#compression]
{quote}
Prefix Tree
Prefix tree encoding was introduced as an experimental feature in HBase 0.96.
It provides similar memory savings to the Prefix, Diff, and Fast Diff encoder,
but provides faster random access at a cost of slower encoding speed.
{quote}
We just want to use random access feature of Prefix Tree. Besides stableness
and slow encoding speed, is there other issues for Prefix tree?
Actually for stableness, I thought it should be stable, since it was introduced
3 years ago.
> PrefixTreeArrayReversibleScanner#previousRowInternal doesn't work correctly
> ---------------------------------------------------------------------------
>
> Key: HBASE-17375
> URL: https://issues.apache.org/jira/browse/HBASE-17375
> Project: HBase
> Issue Type: Bug
> Affects Versions: 1.1.7, 0.98.23, 0.98.24
> Reporter: Chang chen
> Attachments: row trie example.PNG
>
>
> Recently, we find our hbase compaction thread never end. Assume we have
> following cells:
> {quote}
> <A,a> 1
> <A,v> 1
> <Aaeeee,a> 1
> <Aaeeee,v> 1
> <Abc,a> 1
> <Abc,v> 1
> <Abde,a> 1
> <Abde,v> 1
> {quote}
> If we encode above datas into prefix tree block, then it looks like:
> !row trie example.PNG!
> Assume the current row is {color:red}Abc{color} (e.g. the current row node is
> 4), then the previous row should be *Aaeeee* (e.g. 2). However
> previousRowInternal return {color:red}A{color}(e.g. 1)
> After investigation, I believe it's the bug of
> PrefixTreeArrayReversibleScanner#previousRowInternal.
> {code}
> private boolean previousRowInternal() {
> //...
> while (!beforeFirst) {
> //....
> // what if currentRowNode is nub?
> if (currentRowNode.hasOccurrences()) {// escape clause
> currentRowNode.resetFanIndex();
> return true;// found some values
> }
> }
> {code}
> currentRowNode.hasOccurrences() only test whether it has cell or not. But in
> the case of currentRowNode.isNub() is true, previousRowInternal should
> follow the previous fan instead of return.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
