Ummm, you got the gist of it (I may have misspoke in what I initially said).
What my first thought was to make an iterator that will filter down to
the columns that you want. It doesn't look like we have an iterator that
will efficiently do this for you included in the core (although, I know
I've done something similar in the past like this). This iterator would
scan the rows on your table returning just the columns you want.
000200001ccaac30 meta:size [] 1807
000200001cdaac30 meta:size [] 656
000200001cfaac30 meta:size [] 565
Then, we could put the summing combiner on top of that iterator to sum
those and get back a single key. The row in the key you return should be
the last row you included in the sum. This way, if a retry happens under
the hood by the batchscanner, you'll resume where you left off and won't
double-count things.
(you could even do things like sum a maximum of N rows before returning
back some intermediate count to better parallelize things)
000200001cfaac30 meta:size [] 3028
So, each "ScanSession" (what the batchscanner is doing underneath the
hood) would return you a value which your client would do a final summation.
The final stack would be {(data from accumulo) > SKVI to project columns
> summing combiner} > final summation, where {...} denotes work done
server-side. This is one of those things that really shines with the
Accumulo API.
On 3/19/14, 6:40 PM, Russ Weeks wrote:
Hi, Josh,
Thanks very much for your response. I think I get what you're saying,
but it's kind of blowing my mind.
Are you saying that if I first set up an iterator that took my key/value
pairs like,
000200001ccaac30 meta:size [] 1807
000200001ccaac30 meta:source [] data2
000200001cdaac30 meta:filename [] doc02985453
000200001cdaac30 meta:size [] 656
000200001cdaac30 meta:source [] data2
000200001cfaac30 meta:filename [] doc04484522
000200001cfaac30 meta:size [] 565
000200001cfaac30 meta:source [] data2
000200001dcaac30 meta:filename [] doc03342958
And emitted something like,
0 meta:size [] 1807
0 meta:size [] 656
0 meta:size [] 565
And then applied a SummingCombiner at a lower priority than that
iterator, then... it should work, right?
I'll give it a try.
Regards,
-Russ
On Wed, Mar 19, 2014 at 3:33 PM, Josh Elser <[email protected]
<mailto:[email protected]>> wrote:
Russ,
Remember about the distribution of data across multiple nodes in
your cluster by tablet.
A tablet, at the very minimum, will contain one row. Any way to say
that same thing is that a row will never be split across multiple
tablets. The only guarantee you get from Accumulo here is that you
can use a combiner to do you combination across one row.
However, when you combine (pun not intended) another SKVI with the
Combiner, you can do more merging of that intermediate "combined
value" from each row before returning back to the client. You can
think of this approach as doing a multi-level summation.
This still requires one final sum on the client side, but you should
get quite the reduction with this approach over doing the entire sum
client side. You sum the meta:size column in parallel across parts
of the table (server-side) and then client-side you sum the sums
from each part.
I can sketch this out in more detail if it's not clear. HTH
On 3/19/14, 6:18 PM, Russ Weeks wrote:
The accumulo manual states that combiners can be applied to
values which
share the same rowID, column family, and column qualifier. Is
there any
way to adjust this behaviour? I have rows that look like,
000200001ccaac30 meta:size [] 1807
000200001ccaac30 meta:source [] data2
000200001cdaac30 meta:filename [] doc02985453
000200001cdaac30 meta:size [] 656
000200001cdaac30 meta:source [] data2
000200001cfaac30 meta:filename [] doc04484522
000200001cfaac30 meta:size [] 565
000200001cfaac30 meta:source [] data2
000200001dcaac30 meta:filename [] doc03342958
and I'd like to sum up all the values of meta:size across all
rows. I
know I can scan the sizes and sum them on the client side, but I was
hoping there would be a way to do this inside my cluster. Is
mapreduce
my only option here?
Thanks,
-Russ
