Hi,
So I've done this Node-centered accumulator, I've written a small
piece about it :
http://blog.guillaume-pitel.fr/2015/06/spark-trick-shot-node-centered-aggregator/
Hope it can help someone
Guillaume
2015-06-18 15:17 GMT+02:00 Guillaume Pitel guillaume.pi...@exensa.com
Hi,
Thank you for this confirmation.
Coalescing is what we do now. It creates, however, very big partitions.
Guillaume
Hey,
I am not 100% sure but from my understanding accumulators are per partition
(so per task as its the same) and are sent back to the driver with the task
result and
I was thinking exactly the same. I'm going to try it, It doesn't really
matter if I lose an executor, since its sketch will be lost, but then
reexecuted somewhere else.
And anyway, it's an approximate data structure, and what matters are
ratios, not exact values.
I mostly need to take care
Hey,
I am not 100% sure but from my understanding accumulators are per partition
(so per task as its the same) and are sent back to the driver with the task
result and merged. When a task needs to be run n times (multiple rdds
depend on this one, some partition loss later in the chain etc) then
BTW I suggest this instead of using thread locals as I am not sure in which
situation spark will reuse or not them. For example if an error happens
inside a thread, will spark then create a new one or the error is catched
inside the thread preventing it to stop. So in short, does spark guarantee
Yeah thats the problem. There is probably some perfect num of partitions
that provides the best balance between partition size and memory and merge
overhead. Though it's not an ideal solution :(
There could be another way but very hacky... for example if you store one
sketch in a singleton per
Hi,
I'm trying to figure out the smartest way to implement a global
count-min-sketch on accumulators. For now, we are doing that with RDDs.
It works well, but with one sketch per partition, merging takes too long.
As you probably know, a count-min sketch is a big mutable array of array
of
2015-06-18 15:17 GMT+02:00 Guillaume Pitel guillaume.pi...@exensa.com:
I was thinking exactly the same. I'm going to try it, It doesn't really
matter if I lose an executor, since its sketch will be lost, but then
reexecuted somewhere else.
I mean that between the action that will update the