[
https://issues.apache.org/jira/browse/CASSANDRA-1472?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Stu Hood updated CASSANDRA-1472:
--------------------------------
Attachment: v4-bench-c32.txt
v4-bench-c50.txt
Attaching some benchmark numbers generated by CASSANDRA-1531.
{{v4-bench-c32.txt}} is for cardinality == 32, which is a bit of a magic number
in the current implementation of KEYS_BITMAP indexes: it's the maximum number
of bins that our current implementation will create, meaning that every value
gets its own bin, so there are no false positives returned by the sstables.
{{v4-bench-c50.txt}} is for cardinality == 50, meaning that some of the bins
are multi-valued, leading to a larger number of false positives. Since our bin
selection is still naive, it is generating lumpy buckets: you can see a long
tail for KEYS_BITMAP reads, since one thread got stuck with higher cardinality
bins.
----
TODOs:
* Integrate with 1415 now that it is in trunk
* Improve bin selection to eliminate lumpy bins (as seen with c==50)
* Replace Memtable brute-force filtering with an in-memory index attached to
the Memtable (I didn't do any tests with memtables, but I'm assuming this
should still be a priority)
* Implement the "OrBiC" projection described in the paper to eliminate false
positives
** Multi-valued bins would have their values projected into another SSTable
component
> Add bitmap secondary indexes
> ----------------------------
>
> Key: CASSANDRA-1472
> URL: https://issues.apache.org/jira/browse/CASSANDRA-1472
> Project: Cassandra
> Issue Type: Improvement
> Components: Core
> Reporter: Stu Hood
> Fix For: 0.7.0
>
> Attachments: 1472-v3.tgz, 1472-v4.tgz, anatomy.png, v4-bench-c32.txt,
> v4-bench-c50.txt
>
>
> Bitmap indexes are a very efficient structure for dealing with immutable
> data. We can take advantage of the fact that SSTables are immutable by
> attaching them directly to SSTables as a new component (supported by
> CASSANDRA-1471).
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
