[
https://issues.apache.org/jira/browse/PHOENIX-5018?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16720467#comment-16720467
]
Geoffrey Jacoby commented on PHOENIX-5018:
------------------------------------------
Thinking about this some more after talking with [~kozdemir] offline about some
testing he's doing that verified that an UPSERT SELECT into an index does the
right thing and uses the SELECT's KeyValue's timestamps.
While that probably lets non-ASYNC index builds off the hook, I think we still
have a bug with ASYNC and partial rebuilds through the IndexTool. The MapReduce
job runs a SELECT, and each call of map() returns a row into a ResultSet. Those
column values are then put into a JDBC Statement as parameters _into an UPSERT
VALUES_, not an UPSERT SELECT. Since the select and upsert are disconnected, I
don't see how the timestamps could be connected since the UPSERT never sees the
original KeyValues.
Easiest way to verify this would probably be adding tests to IndexToolIT that
assert the rebuilt index can still be seen with the same SCN that the original
data had.
> Index mutations created by IndexTool will have wrong timestamps
> ---------------------------------------------------------------
>
> Key: PHOENIX-5018
> URL: https://issues.apache.org/jira/browse/PHOENIX-5018
> Project: Phoenix
> Issue Type: Bug
> Affects Versions: 4.14.0, 5.0.0
> Reporter: Geoffrey Jacoby
> Assignee: Kadir OZDEMIR
> Priority: Major
>
> When doing a full rebuild (or initial async build) on an index using the
> IndexTool and PhoenixIndexImportDirectMapper, we generate the index mutations
> by creating an UPSERT SELECT query from the base table to the index, then
> taking the Mutations from it and inserting it directly into the index via an
> HBase HTable.
> The timestamps of the Mutations use the default HBase behavior, which is to
> take the current wall clock. However, the timestamp of an index KeyValue
> should use the timestamp of the initial KeyValue in the base table.
> Having base table and index timestamps out of sync can cause all sorts of
> weird side effects, such as if the base table has data with an expired TTL
> that isn't expired in the index yet.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
