Thank you Aaron, this was very helpful.
Could it be an issue that this optimization does not really take effect
until the memtable with the hoisted data is flushed? In my simple
example below, the same row is updated and multiple selects of the same
row will result in multiple writes to the memtable. It seems it maybe
possible (although unlikely) where, if you go from a write-mostly to a
read-mostly scenario, you could get into a state where you are stuck
rewriting to the same memtable, and the memtable is not flushed because
it absorbs the over-writes. I can foresee this especially if you are
reading the same rows repeatedly.
I also noticed from the codepaths that if Row caching is enabled, this
optimization will not occur. We made some changes this weekend to make
this column family more suitable to row-caching and enabled row-caching
with a small cache. Our initial results is that it seems to have
corrected the write counts, and has increased performance quite a bit.
However, are there any hidden gotcha's there because this optimization
is not occurring? https://issues.apache.org/jira/browse/CASSANDRA-2503
mentions a "compaction is behind" problem. Any history on that? I
couldn't find too much information on it.
Thanks,
-Mike
On 12/16/2012 8:41 PM, aaron morton wrote:
1) Am I reading things correctly?
Yes.
If you do a read/slice by name and more than min compaction level
nodes where read the data is re-written so that the next read uses
fewer SSTables.
2) What is really happening here? Essentially minor compactions can
occur between 4 and 32 memtable flushes. Looking through the code,
this seems to only effect a couple types of select statements (when
selecting a specific column on a specific key being one of them).
During the time between these two values, every "select" statement
will perform a write.
Yup, only for readying a row where the column names are specified.
Remember minor compaction when using SizedTiered Compaction (the
default) works on buckets of the same size.
Imagine a row that had been around for a while and had fragments in
more than Min Compaction Threshold sstables. Say it is 3 SSTables in
the 2nd tier and 2 sstables in the 1st. So it takes (potentially) 5
SSTable reads. If this row is read it will get hoisted back up.
But the row has is in only 1 SSTable in the 2nd tier and 2 in the 1st
tier it will not hoisted.
There are a few short circuits in the SliceByName read path. One of
them is to end the search when we know that no other SSTables contain
columns that should be considered. So if the 4 columns you read
frequently are hoisted into the 1st bucket your reads will get handled
by that one bucket.
It's not every select. Just those that touched more the min compaction
sstables.
3) Is this desired behavior? Is there something else I should be
looking at that could be causing this behavior?
Yes.
https://issues.apache.org/jira/browse/CASSANDRA-2503
Cheers
-----------------
Aaron Morton
Freelance Cassandra Developer
New Zealand
@aaronmorton
http://www.thelastpickle.com
On 15/12/2012, at 12:58 PM, Michael Theroux <mthero...@yahoo.com
<mailto:mthero...@yahoo.com>> wrote:
Hello,
We have an unusual situation that I believe I've reproduced, at least
temporarily, in a test environment. I also think I see where this
issue is occurring in the code.
We have a specific column family that is under heavy read and write
load on a nightly basis. For the purposes of this description, I'll
refer to this column family as "Bob". During this nightly
processing, sometimes Bob is under very write load, other times it is
very heavy read load.
The application is such that when something is written to Bob, a
write is made to one of two other tables. We've witnessed a
situation where the write count on Bob far outstrips the write count
on either of the other tables, by a factor of 3->10. This is based
on the WriteCount available on the column family JMX MBean. We have
not been able to find where in our code this is happening, and we
have gone as far as tracing our CQL calls to determine that the
relationship between Bob and the other tables are what we expect.
I brought up a test node to experiment, and see a situation where,
when a "select" statement is executed, a write will occur.
In my test, I perform the following (switching between nodetool and
cqlsh):
update bob set 'about'='coworker' where key='<hex key>';
nodetool flush
update bob set 'about'='coworker' where key='<hex key>';
nodetool flush
update bob set 'about'='coworker' where key='<hex key>';
nodetool flush
update bob set 'about'='coworker' where key='<hex key>';
nodetool flush
update bob set 'about'='coworker' where key='<hex key>';
nodetool flush
Then, for a period of time (before a minor compaction occurs), a
select statement that selects specific columns will cause writes to
occur in the write count of the column family:
select about,changed,data from bob where key='<hex key>';
This situation will continue until a minor compaction is completed.
I went into the code and added some traces to CollationController.java:
private ColumnFamily collectTimeOrderedData()
{
logger.debug("collectTimeOrderedData");
... <snip> ...
---> HERE logger.debug( "tables iterated: " + sstablesIterated + " Min compact:
" + cfs.getMinimumCompactionThreshold() );
// "hoist up" the requested data into a more recent sstable
if (sstablesIterated > cfs.getMinimumCompactionThreshold()
&& !cfs.isCompactionDisabled()
&& cfs.getCompactionStrategy() instanceof
SizeTieredCompactionStrategy)
{
RowMutation rm = new RowMutation(cfs.table.name, new
Row(filter.key, returnCF.cloneMe()));
try
{
---> HERE logger.debug( "Apply hoisted up row mutation" );
// skipping commitlog and index updates is fine since
we're just de-fragmenting existing data
Table.open(rm.getTable()).apply(rm, false, false);
}
catch (IOException e)
{
// log and allow the result to be returned
logger.error("Error re-writing read results", e);
}
}
... <snip> ...
Performing the steps above, I see the following traces (in the test
environment I decreased the minimum compaction threshold to make this
easier to reproduce). After I do a couple of update/flush, I see this
in the log:
DEBUG [FlushWriter:7] 2012-12-14 22:54:40,106 CompactionManager.java
(line 117) Scheduling a background task check for bob with
SizeTieredCompactionStrategy
Then, until compaction occurs, I see (when performing a select):
DEBUG [ScheduledTasks:1] 2012-12-14 22:55:15,998 LoadBroadcaster.java
(line 86) Disseminating load info ...
DEBUG [Thrift:12] 2012-12-14 22:55:16,990 CassandraServer.java (line
1227) execute_cql_query
DEBUG [Thrift:12] 2012-12-14 22:55:16,991 QueryProcessor.java (line
445) CQL statement type: SELECT
DEBUG [Thrift:12] 2012-12-14 22:55:16,991 StorageProxy.java (line
653) Command/ConsistencyLevel is
SliceByNamesReadCommand(table='open',
key=804229d1933669d0a25d2a38c8b26ded10069573003e6dbb1ce21b5f402a5342,
columnParent='QueryPath(columnFamilyName='bob',
superColumnName='null', columnName='null')',
columns=[about,changed,data,])/ONE
DEBUG [Thrift:12] 2012-12-14 22:55:16,992 ReadCallback.java (line 79)
Blockfor is 1; setting up requests to /10.0.4.20
DEBUG [Thrift:12] 2012-12-14 22:55:16,992 StorageProxy.java (line
669) reading data locally
DEBUG [ReadStage:61] 2012-12-14 22:55:16,992 StorageProxy.java (line
813) LocalReadRunnable reading SliceByNamesReadCommand(table='open',
key=804229d1933669d0a25d2a38c8b26ded10069573003e6dbb1ce21b5f402a5342,
columnParent='QueryPath(columnFamilyName='bob',
superColumnName='null', columnName='null')',
columns=[about,changed,data,])
DEBUG [ReadStage:61] 2012-12-14 22:55:16,992 CollationController.java
(line 68) In get top level columns: class
org.apache.cassandra.db.filter.NamesQueryFilter type: Standard valid:
class org.apache.cassandra.db.marshal.BytesType
DEBUG [ReadStage:61] 2012-12-14 22:55:16,992 CollationController.java
(line 84) collectTimeOrderedData
---> DEBUG [ReadStage:61] 2012-12-14 22:55:17,192
CollationController.java (line 188) tables iterated: 4 Min compact: 2
----> DEBUG [ReadStage:61] 2012-12-14 22:55:17,192
CollationController.java (line 198) Apply hoisted up row mutation
DEBUG [ReadStage:61] 2012-12-14 22:55:17,193 Table.java (line 395)
applying mutation of row
804229d1933669d0a25d2a38c8b26ded10069573003e6dbb1ce21b5f402a5342
The above traces will occur every time I repeat the above select
statement.
Minor compaction doesn't start until a few minutes after the request
was submitted above (note, this is an unloaded test node):
DEBUG [CompactionExecutor:11] 2012-12-14 22:57:03,278
IntervalNode.java (line 45) Creating IntervalNode from
[Interval(DecoratedKey(Token(bytes[804229d1933669d0a25d2a38c8b26ded10069573003e6dbb1ce...
Once minor compaction occurs, the behavior around write count being
incremented stops, until more than the minimum compaction threshold
memtables are flush to disk.
So, my questions are:
1) Am I reading things correctly?
2) What is really happening here? Essentially minor compactions can
occur between 4 and 32 memtable flushes. Looking through the code,
this seems to only effect a couple types of select statements (when
selecting a specific column on a specific key being one of them).
During the time between these two values, every "select" statement
will perform a write.
3) Is this desired behavior? Is there something else I should be
looking at that could be causing this behavior?
We are running Cassandra 1.1.2, with SizeTieredCompactionStrategy.
Any help is appreciated,
Thanks,
-Mike
