[
https://issues.apache.org/jira/browse/HBASE-15398?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15203828#comment-15203828
]
Phil Yang commented on HBASE-15398:
-----------------------------------
Before we support partial protocol, we already have family essential filters,
they mainly used for reduce the time of scanning if we skip the row because we
won't read any data in joinedHeap. It also works now, right? So at least we
won't make anything worse.
Let me summarize what we have:
If we don't want to break the assumption that client only receives sorted
cells, we have to disable partial protocol if the order of partial results may
be disorder.
Based on this:
(1)Before this issue we allow user using family essential filters and
hasFilterRow return false(it must be a user-defined filter) at the same time.
But we will ban it after this issue resolved because we can not use partial
protocol here and we don't want to increase the probability of OOM/timeout.
(2)If user using family essential filters and hasFilterRow return true(eg.
SCVF), before this issue the order of results is wrong when using partial
protocol, after this issue resolved we will disable partial protocol. So here
we will increase the probability of OOM/timeout.
(3)In 1.1.0-1.1.3 and 1.2.0, we allow user using setAllowPartial(true) and
family essential filters at the same time. But the order of results is wrong.
After this issue resolved we will ban this usage because we must disable
partial protocol.
Any supplement or concerns?
> Cells loss or disorder when using family essential filter and partial
> scanning protocol
> ---------------------------------------------------------------------------------------
>
> Key: HBASE-15398
> URL: https://issues.apache.org/jira/browse/HBASE-15398
> Project: HBase
> Issue Type: Bug
> Components: dataloss, Scanners
> Affects Versions: 1.2.0, 1.1.3
> Reporter: Phil Yang
> Assignee: Phil Yang
> Priority: Critical
> Attachments: 15398-test.txt, HBASE-15398-v2.patch,
> HBASE-15398-v3.patch, HBASE-15398-v4.patch, HBASE-15398-v5.patch,
> HBASE-15398.v1.txt
>
>
> In RegionScannerImpl, we have two heaps, storeHeap and joinedHeap. If we have
> a filter and it doesn't apply to all cf, the stores whose families needn't be
> filtered will be in joinedHeap. We scan storeHeap first, then joinedHeap,
> and merge the results and sort and return to client. We need sort because the
> order of Cell is rowkey/cf/cq/ts and a smaller cf may be in the joinedHeap.
> However, after HBASE-11544 we may transfer partial results when we get
> SIZE_LIMIT_REACHED_MID_ROW or other similar states. We may return a larger cf
> first because it is in storeHeap and then a smaller cf because it is in
> joinedHeap. Server won't hold all cells in a row and client doesn't have a
> sorting logic. The order of cf in Result for user is wrong.
> And a more critical bug is, if we get a LIMIT_REACHED_MID_ROW on the last
> cell of a row in storeHeap, we will break scanning in RegionScannerImpl and
> in populateResult we will change the state to SIZE_LIMIT_REACHED because next
> peeked cell is next row. But this is only the last cell of one and we have
> two... And SIZE_LIMIT_REACHED means this Result is not partial (by
> ScannerContext.partialResultFormed), client will see it and merge them and
> return to user with losing data of joinedHeap. On next scan we will read next
> row of storeHeap and joinedHeap is forgotten and never be read...
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
