Quote Nathan's comments in
https://groups.google.com/d/msg/storm-user/xMt3q_jJeZA/jJkZm3TtsZUJ
There is overhead to each batch of tuples that Trident processes. So if you
> only have a handful of tuples in each batch, the overhead will dominate (in
> addition to serialization that Ted brought up). In normal operation – on a
> cluster with non-trivial batches – Trident's performance is excellent. I
> benchmarked it as having essentially the same performance as vanilla Storm.
Seems Trident is tightly binded with concept of batch, which will introduce
performance overhead. Some user don't care about transaction, but want to
use Trident to express and optimize the topology.
|