cassandra-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Daniel Doubleday (JIRA)" <>
Subject [jira] [Commented] (CASSANDRA-2864) Alternative Row Cache Implementation
Date Mon, 25 Jun 2012 10:22:44 GMT


Daniel Doubleday commented on CASSANDRA-2864:

So - same here - so mutch to do, so little time ...

bq. The serialization format for columns seems only marginaly different from our internal
one. Maybe it would be worth reusing ColumnSerializer

Yes I guess it would make the code a bit cleaner. CachedRow could become a DataInput and BBU
can special case read(). The main thing is that searching should not copy bytes or construct
columns while comparing column names. At least my test showed my that this is where I lost
most of the performance. You will add a byte or so in serialized size because the column index
needs to be fixed width and offset information must be an int and is redundant if you use
standard serialization.

bq. it would make sense to allow serializing off-heap

I thought about that but did not experiment. Actually I'm not so sure that it will make sense
because the main idea of off-heap serialization here is reducing GC pressure. But the point
is that you wont have less objects in this case. So it would only help when the ByteBuffers
cause fragmentation problems. Since malloc is not a silver bullet either well ... But it should
be so ease to implement that experimenting wouldn't hurt

bq. What is the point of collectTimeOrderedData in RowCacheCollationController

I had a test with many writes which showed that the overhead of testing the cached row for
a column that is superseded by a memtable value is significant and I wanted to avoid that
in this case. And at that point I still hoped that I can support counters out of the box.

bq. What's the goal of noMergeNecessary


bq. instead of having two collation controllers

Definitely. It was just easier to maintain for us as a patch

Now: I updated the patch.

- Should apply cleanly on trunk
- Support for CASSANDRA-3885 (untested)
- No support counters
- No support for CASSANDRA-3708

I tried to update the patch for CASSANDRA-3708 but failed because I couldn't get a firm understanding
in the limited time. I.e. it seems that name queries will not return a range tomb stone [1..3]
when asked for column 2. Also a range tomb stone [1..3] seems to overwrite [1..4]. Both seems
strange but might be a misunderstanding or it's supposed to work that way.

In short: I gave up on this one for the time being.

My time is pretty limited unfortunately so Sylvain if you want to take it from here ...
Otherwise I would need some support regarding CASSANDRA-3708.

> Alternative Row Cache Implementation
> ------------------------------------
>                 Key: CASSANDRA-2864
>                 URL:
>             Project: Cassandra
>          Issue Type: Improvement
>          Components: Core
>            Reporter: Daniel Doubleday
>            Assignee: Daniel Doubleday
>              Labels: cache
>             Fix For: 1.2
>         Attachments: 0001-CASSANDRA-2864-w-out-direct-counter-support.patch
> we have been working on an alternative implementation to the existing row cache(s)
> We have 2 main goals:
> - Decrease memory -> get more rows in the cache without suffering a huge performance
> - Reduce gc pressure
> This sounds a lot like we should be using the new serializing cache in 0.8. 
> Unfortunately our workload consists of loads of updates which would invalidate the cache
all the time.
> *Note: Updated Patch Description (Please check history if you're interested where this
was comming from)*
> h3. Rough Idea
> - Keep serialized row (ByteBuffer) in mem which represents unfiltered but collated columns
of all ssts but not memtable columns
> - Writes dont affect the cache at all. They go only to the memtables
> - Reads collect columns from memtables and row cache
> - Serialized Row is re-written (merged) with mem tables when flushed
> h3. Some Implementation Details
> h4. Reads
> - Basically the read logic differ from regular uncached reads only in that a special
CollationController which is deserializing columns from in memory bytes
> - In the first version of this cache the serialized in memory format was the same as
the fs format but test showed that performance sufferd because a lot of unnecessary deserialization
takes place and that columns seeks are O( n ) whithin one block
> - To improve on that a different in memory format was used. It splits length meta info
and data of columns so that the names can be binary searched. 
> {noformat}
> ===========================
> Header (24)                    
> ===========================
> MaxTimestamp:        long  
> LocalDeletionTime:   int   
> MarkedForDeleteAt:   long  
> NumColumns:          int   
> ===========================
> Column Index (num cols * 12)              
> ===========================
> NameOffset:          int   
> ValueOffset:         int   
> ValueLength:         int   
> ===========================
> Column Data                
> ===========================
> Name:                byte[]
> Value:               byte[]
> SerializationFlags:  byte  
> Misc:                ?     
> Timestamp:           long  
> ---------------------------
> Misc Counter Column        
> ---------------------------
> TSOfLastDelete:      long  
> ---------------------------
> Misc Expiring Column       
> ---------------------------
> TimeToLive:          int   
> LocalDeletionTime:   int   
> ===========================
> {noformat}
> - These rows are read by 2 new column interators which correspond to SSTableNamesIterator
and SSTableSliceIterator. During filtering only columns that actually match are constructed.
The searching / skipping is performed on the raw ByteBuffer and does not create any objects.
> - A special CollationController is used to access and collate via cache and said new
iterators. It also supports skipping the cached row by max update timestamp
> h4. Writes
> - Writes dont update or invalidate the cache.
> - In CFS.replaceFlushed memtables are merged before the data view is switched. I fear
that this is killing counters because they would be overcounted but my understading of counters
is somewhere between weak and non-existing. I guess that counters if one wants to support
them here would need an additional unique local identifier in memory and in serialized cache
to be able to filter duplicates or something like that.
> {noformat}
>     void replaceFlushed(Memtable memtable, SSTableReader sstable)
>     {
>         if (sstCache.getCapacity() > 0) {
>             mergeSSTCache(memtable);
>         }
>         data.replaceFlushed(memtable, sstable);
>         CompactionManager.instance.submitBackground(this);
>     }
> {noformat}
> Test Results: See comments below

This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:!default.jspa
For more information on JIRA, see:


View raw message