accumulo-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From William Slacum <>
Subject Re: Iterating/Aggregating/Combining Complex (Java POJO/Avro) Values
Date Mon, 14 Jul 2014 19:44:55 GMT
Hi Mike!

The Combiner interface is only for aggregating keys within a single row.
You can probably get away with implementing your combining logic in a
WrappingIterator that reads across all the rows in a given tablet.

To do some combine/fold/reduce operation, Accumulo needs the input type to
be the same as the output type. The combiner doesn't have a notion of a
"present" type (as you'd see in something like Algebird's Groups), but you
can use another iterator to perform your transformation.

If you wanted to extract the "count" field from your Avro object, you could
write a new Iterator that took your Avro object, extracted the desired
field, and returned it as its top value. You can then set this iterator as
the source of the aggregator, either programmatically or via by wrapping
the source object passed to the aggregator in its
SortedKeyValueIterator#init call.

This is a bit inefficient as you'd have to serialize to a Value and then
immediately deserialize it in the iterator above it. You could mitigate
this by exposing a method that would get the extracted value before
serializing it.

This kind of counting also requires client side logic to do a final combine
operation, since the aggregations from all the tservers are partial results.

I believe that CountingIterator is not meant for user consumption, but I do
not know if it's related to your issue in trying to use it from the shell.
Iterators set through the shell, in previous versions of Accumulo, have a
requirement to implement OptionDescriber. Many default iterators do not
implement this, and thus can't set in the shell.

On Mon, Jul 14, 2014 at 2:44 PM, Michael Moss <>

> Hi, All.
> I'm curious what the best practices are around persisting complex
> types/data in Accumulo (and aggregating on fields within them).
> Let's say I have (row, column family, column qualifier, value):
> "A" "foo" "" MyHugeAvroObject(count=2)
> "A" "foo" "" MyHugeAvroObject(count=3)
> Let's say MyHugeAvroObject has a field "Integer count" with the values
> above.
> What is the best way to aggregate on row, column family, column qualifier
> by count? In my above example:
> "A" "foo" "" 5
> The TypedValueCombiner.typedReduce method can deserialize any "V", in my
> case MyHugeAvroObject, but it needs to return a value of type "V". What are
> the best practices for deeply nested/complex objects? It's not always
> straightforward to map a complex Avro type into Row -> Column Family ->
> Column Qualifier.
> Rather than using a TypedCombiner, I looked into using an Aggregator
> (which appears deprecated as of 1.4), which appears to let me return
> arbitrary values, but despite running setiter, my aggregator doesn't seem
> to do anything.
> I also tried looking at implementing a WrappingIterator, which also
> appears to allow me to return arbitary values (such as Accumulo's
> CountingIterator), but I get cryptic errors when trying to setiter, I'm on
> Accumulo 1.6:
> root@dev kyt> setiter -t kyt -scan -p 10 -n countingIter -class
> org.apache.accumulo.core.iterators.system.CountingIterator
> 2014-07-14 11:12:55,623 [shell.Shell] ERROR:
> java.lang.IllegalArgumentException:
> org.apache.accumulo.core.iterators.system.CountingIterator
> This is odd because other included implementations of WrappingIterator
> seem to work (perhaps the implementation of CountingIterator is dated):
> root@dev kyt> setiter -t kyt -scan -p 10 -n deletingIterator -class
> org.apache.accumulo.core.iterators.system.DeletingIterator
> The iterator class does not implement OptionDescriber. Consider this for
> better iterator configuration using this setiter command.
> Name for iterator (enter to skip):
> All in all, how can I aggregate simple values, like counters from rows
> with complex Avro objects as Values without having to add aggregations
> fields to these Value objects?
> Thanks!
> -Mike

View raw message