[
https://issues.apache.org/jira/browse/CASSANDRA-10681?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15011664#comment-15011664
]
Pavel Yaskevich commented on CASSANDRA-10681:
---------------------------------------------
Let me clarify a bit - it serializes a compilation of all of the indexes not
building them. That's what I have already mentioned, Indexes are built separate
but it's just a nature of the Index API since PerRowIndex is no more we have to
build all of the indexes independently, this requires to pass through a data
multiple times but I don't think it's necessary a problem if we list the
assumption that indexes are build in one and only way - by merging sstables
together and feeding index collated row - and let API implementers decide how
to build indexes based on the set of sstables. When the SSTable is added via
streaming for example CASSANDRA-10678 would take care of creating indexes for
it in case of SASI and Indexer API in case of standard indexes, so I don't
really see a problem there, in case of side loading new compaction task per
index is going to be triggered to build such indexes in necessary but we can't
really go around that.
> make index building pluggable via IndexBuildTask
> ------------------------------------------------
>
> Key: CASSANDRA-10681
> URL: https://issues.apache.org/jira/browse/CASSANDRA-10681
> Project: Cassandra
> Issue Type: Sub-task
> Components: Local Write-Read Paths
> Reporter: Pavel Yaskevich
> Assignee: Pavel Yaskevich
> Priority: Minor
> Labels: sasi
> Fix For: 3.x
>
>
> Currently index building assumes one and only way to build all of the indexes
> - through SecondaryIndexBuilder - which merges all of the sstables together,
> collates columns etc. Such works fine for built-in indexes but not for SASI
> since it's attaches to every SSTable individually. We need a "IndexBuildTask"
> interface (based on CompactionInfo.Holder) to be returned from Index on
> demand to give power to SI interface implementers to decide how build should
> work. This might be less effective for CassandraIndex, since this effectively
> means that collation will have to be done multiple times on the same data,
> but nevertheless is a good compromise for clean interface to outside world.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
