cassandra-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Ross M (JIRA)" <>
Subject [jira] Commented: (CASSANDRA-836) CommitLogSegment::seekAndWriteCommitLogHeader assumes header size doesn't change.
Date Sun, 28 Feb 2010 21:16:26 GMT


Ross M commented on CASSANDRA-836:

i'm not a moron. that was an example attached to the bug that quickly caused the problem to

that said it actually is smaller than the java serialization of the BitSet {2, 4} of length
5. with ints 5, 2, 2, 4 stored (128 bits) vs the 
 "java.util.BitSet" portion of java serialization which takes up 128 bits then you have at
least 64-bits for the actual bit values plus anything else it sticks in there. there apparently
are other things b/c the default version results in ~592 bits of data. even with all 5 bits
turned on it would take 224 with the int version.

> CommitLogSegment::seekAndWriteCommitLogHeader assumes header size doesn't change.
> ---------------------------------------------------------------------------------
>                 Key: CASSANDRA-836
>                 URL:
>             Project: Cassandra
>          Issue Type: Bug
>          Components: Core
>         Environment: n/a - all
>            Reporter: Ross M
>            Priority: Minor
>         Attachments:
> CommitLogSegment::seekAndWriteCommitLogHeader assumes header size doesn't grow. there
are pieces of the header (BitSet) that are serialized with java serialization which makes
no such promises. 
> the following code:
>     /** writes header at the beginning of the file, then seeks back to current position
>     void seekAndWriteCommitLogHeader(byte[] bytes) throws IOException
>     {
>         long currentPos = logWriter.getFilePointer();
>         writeCommitLogHeader(bytes);
>     }
> works fine as long as the header size doesn't change, but if it grows the new header
will over write the beginning of the data segment. the bit-set being written in the header
happens to serialize to the same size, but there is no guarantee of this.
> i found this when looking at optimizing the serialization of data to disk (thus improving
write throughput/performance.) i removed the ObjectOutputStream serialization in BitSetSerializer
and replaced it with a custom serialization that omits the generic java serialization/ObjectOutputStream
stuff and just writes on the "true" bits. the custom serialization worked fine, but broke
other parts of the code when the header bitset had new bits turned on, thus growing the header's
size, data segment bytes were overwritten.
> the serialized version of a BitSet can grow in a similar manner, no pomises of size/consistency
are made, but with current use it luckily doesn't seem to happen.
> a good fix is unclear. without forcing the header to be a fixed/constant size in some
manner this problem could pop up at any point. it's generally not safe to rewrite headers
like this without custom code that ensures the size doesn't change. one fix would be to manually
write all of the header data out (rather than relying on java serialization and serialization
code in other parts of cassandra not to change.) another might be to pad the size of the header
so that the data inside can grow, but that seems fraught with (potential) problems. (i've
played around with padding the header length, but that seems to cause other things to break,
which i haven't been able to track down yet.)

This message is automatically generated by JIRA.
You can reply to this email to add a comment to the issue online.

View raw message