[
https://issues.apache.org/jira/browse/ACCUMULO-652?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13409808#comment-13409808
]
Adam Fuchs commented on ACCUMULO-652:
-------------------------------------
We uncovered another tricky point today: if we use timestamp ranges to filter
out blocks that contain deletes, we might re-introduce entries that have been
deleted. This would break the established semantics that support guaranteeing a
logical, irreversible purge of entries (almost as bad as crossing the streams).
In the same family of problems, a TimestampRangeFilter iterator would not be
commutative with the VersioningIterator or any Aggregator because it would lead
to incomplete/inconsistent results.
In the delete case, we need to add an index block stat that keeps track of the
greatest timestamp of any delete entry. Then when we do the filtering we can
include any blocks that might have deletes that are greater than the minimum
timestamp in the timestamp range. This is a "must do".
The second case is a bit trickier. In the general case, we need to pull back
any blocks whose key range includes any of the keys that match the given
timestamp range. In the VersioningIterator case, an alternative solution would
be to extend the timestamp range to include anything greater than the minimum
timestamp, ignoring the max timestamp. For now, I think we need to punt on the
general case and just say that the TimestampRangeFilter and other
versioning/aggregation iterators are simply not compatible in all caps in the
javadocs. Longer term, this should be an exemplar when we rewrite the iterator
framework.
> support block-based filtering within RFile
> ------------------------------------------
>
> Key: ACCUMULO-652
> URL: https://issues.apache.org/jira/browse/ACCUMULO-652
> Project: Accumulo
> Issue Type: Bug
> Reporter: Adam Fuchs
> Assignee: Adam Fuchs
>
> If we keep some stats about what is in an RFile block, we might be able to
> efficiently [O(log N)], with high probability, implement filters that
> currently require linear table scans. Two use cases of this include timestamp
> range filtering (i.e. give me everything from last Tuesday) and cell-level
> security filtering (i.e. give me everything that I can see with my
> authorizations).
> For the timestamp range filter, we can keep minimum and maximum timestamps
> across all keys used in a block within the index entry for that block. For
> the cell-level security filter, we can keep an aggregate label. This could be
> done using a simplified disjunction of all of the labels in the block. The
> extra block statistics information can propagate up the index hierarchy as
> well, giving nice performance characteristics for finding the next matching
> entry in a file.
> In general, this is a heuristic technique that is good if data tends to
> naturally cluster in blocks with respect to the way it is queried. Testing
> its efficacy will require closely emulating real-world use cases -- tests
> like the continuous ingest test will not be sufficient. We will have to test
> for a few things:
> # The cost for storing the extra stats in the index are not too expensive.
> # The performance benefit for common use cases is significant.
> # We shouldn't introduce any unacceptable worst-case behavior, like bloating
> the index to ridiculous proportions for any data set.
> Eventually this will all need to be exposed through the Iterator API to be
> useful, which will be another ticket.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
