[
https://issues.apache.org/jira/browse/CASSANDRA-790?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12833927#action_12833927
]
Jonathan Ellis commented on CASSANDRA-790:
------------------------------------------
patch 01 comments:
refactoring:
- please split the refactoring into a separate patch; it's hard to tell what
is part of the actual fix here
- BF constructors that do not chain is a design smell; one of them only being
called from tests is also a smell
- instead of using min/max to force values into acceptable ranges, assert that
they are sane
- I feel part of the BF problems here is that BF is trying to be too
high-level. Wouldn't we be better served by having a low-level BF constructor
taking hash & bucket counts, and then factories to do the high level things?
fix:
- this feels like we're trading an obvious problem (BF constructor throws) for
a more subtle one (BF is a no-op when we exceed the spec, as noted by TODO).
wouldn't it be better to log a warning, create the largest BF possible, and
degrade gracefully? This would be easier if the BF constructor were sane as
mentioned above.
> SSTables limited to (2^31)/15 keys
> ----------------------------------
>
> Key: CASSANDRA-790
> URL: https://issues.apache.org/jira/browse/CASSANDRA-790
> Project: Cassandra
> Issue Type: Bug
> Affects Versions: 0.5, 0.6, 0.7
> Reporter: Stu Hood
> Priority: Blocker
> Fix For: 0.5, 0.6, 0.7
>
> Attachments:
> 0001-Change-parameters-to-BloomCalculations-in-order-to-c.patch,
> 0002-Add-timeouts-to-forceBlockingFlush-during-tests.patch
>
>
> The current BloomFilter implementation requires a BitSet of (bucket_count *
> num_keys) in size, and that calculation is currently performed in an integer,
> which causes overflow for around 140 million keys in one SSTable.
> Short term fix: perform the calculation in a long, and cap the value to the
> maximum size of a BitSet.
> Long term fix: begin partitioning BitSets, perhaps using Linear Bloom Filters.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
