[
https://issues.apache.org/jira/browse/CASSANDRA-5417?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13621017#comment-13621017
]
Sylvain Lebresne commented on CASSANDRA-5417:
---------------------------------------------
I've pushed some initial patch to
https://github.com/pcmanus/cassandra/commits/5417. It's on top of
CASSANDRA-5125 as said above (so only the 3 last commits are for this ticket).
I'll note right away that this patch mainly concern itself with the ground
work, not with the different optimization I discuss in the description.
Typically, it doesn't do any prefix sharing and a bit more work will be
necessary to get that, but it makes it at least possible. Also, while the patch
does clean up the CQL3 code quite a bit imo, I think there is more cleanups
opportunity that are not taken just yet, but I feel that the patch is huge
enough like that, so I'd rather keep that for later.
The patch passes the unit tests as well as (almost all) the cql dtests. For the
cql dtests, due to some refactor, the patch introduce a small limitation that
if you use a IN on the last clustering key, you have to include that last
clustering key in the selected columns. This makes one test fail. We should
definitively fix it, but as it happens, the solution for that is exactly the
one needed for CASSANDRA-4911. So since that's a bit of a detail, and since
this patch is already pretty big, I suggest leaving the fix of that remaining
dtest to CASSANDRA-4911 and I'll take on myself to fix that latter ticket on
2.0 timeframe.
I'll also note that this patch does bias the code a bit towards Composite, in
the sense that for non-composite CF, we do allocate one more object per column
(compare to without the patch). I haven't done much performance testing at that
point however.
> Push composites support in the storage engine
> ---------------------------------------------
>
> Key: CASSANDRA-5417
> URL: https://issues.apache.org/jira/browse/CASSANDRA-5417
> Project: Cassandra
> Issue Type: Improvement
> Reporter: Sylvain Lebresne
> Assignee: Sylvain Lebresne
> Fix For: 2.0
>
>
> CompositeType happens to be very useful and is now widely used: CQL3 heavily
> rely on it, and super columns are now using it too internally. Besides,
> CompositeType has been advised as a replacement of super columns on the
> thrift side for a while, so it's safe to assume that it's generally used
> there too.
> CompositeType has initially been introduced as just another AbstractType.
> Meaning that the storage engine has no nothing whatsoever of composites
> being, well, composite. This has the following drawbacks:
> * Because internally a composite value is handled as just a ByteBuffer, we
> end up doing a lot of extra work. Typically, each time we compare 2 composite
> value, we end up "deserializing" the components (which, while it doesn't copy
> data per-se because we just slice the global ByteBuffer, still waste some cpu
> cycles and allocate a bunch of ByteBuffer objects). And since compare can be
> called *a lot*, this is likely not negligible.
> * This make CQL3 code uglier than necessary. Basically, CQL3 makes extensive
> use of composites, and since it gets backs ByteBuffer from the internal
> columns, it always have to check if it's actually a compositeType or not, and
> then split it and pick the different parts it needs. It's only an API
> problem, but having things exposed as composites directly would definitively
> make thinks cleaner. In particular, in most cases, CQL3 don't care whether it
> has a composite with only one component or a non-really-composite value, but
> we still always distinguishes both cases. Lastly, if we do expose composites
> more directly internally, it's not a lot more work to "internalize" better
> the different parts of the cell name that CQL3 uses (what's the clustering
> key, what's the actuall CQL3 column name, what's the collection element),
> making things cleaner. Last but not least, there is currently a bunch of
> places where methods take a ByteBuffer as argument and it's hard to know
> whether it expects a cell name or a CQL3 column name. This is pretty error
> prone.
> * It makes it hard (or impossible) to do a number of performance
> improvements. Consider CASSANDRA-4175, I'm not really sure how you can do it
> properly (in memory) if cell names are just ByteBuffer (since CQL3 column
> names are just one of the component in general). But we also miss
> oportunities of sharing prefixes. If we were able to share prefixes of
> composite names in memory we would 1) lower the memory footprint and 2)
> potentially speed-up comparison (of the prefixes) by checking reference
> equality first (also, doing prefix sharing on-disk, which is a separate
> concern btw, might be easier to do if we do prefix sharing in memory).
> So I suggest pushing CompositeType support inside the storage engine. What I
> mean by that concretely would be change the internal {{Column.name}} from
> ByteBuffer to some CellName type. A CellName would API-wise just be a list of
> ByteBuffer. But in practice, we'd have a specific CellName implementation for
> not-really-composite names, and the truly composite implementation will allow
> some prefix sharing. From an external API however, nothing would change, we
> would pack the composite as usual before sending it back to the client, but
> at least internally, comparison won't have to deserialize the components
> every time, and CQL3 code will be cleaner.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
