Hi Huangrui,
You can change the code in stats.clj to decrease the statistical
interval. In line 164, you can set the size for stat buckets. The
original value for the smallest bucket is 30, which means 20second * 30
= 10min. You can set the value to be 3, if you want smallest interval to
be 1 min.
I am sorry that I can't answer your second question.
In the current implementation, when an executor is reassigned to a new
slot, say another worker, both the original worker and the target worker
will be killed and restarted. Actually, we are working on the feature of
live executor migration, which allows migrating an executor from a
worker to another one without killing the workers. We plan to submit the
pull request in the near future.
Hope that helps.
Regards,
Li Wang
On 16/12/2015 21:44, R H wrote:
Hi, I have three questions.
In Storm UI the Complete Latency(namely 11.47ms in the pic) is calculated from
the metrics from the last ten minutes. I am trying some scheduling algorithms
which needs faster respond in the measurement.
I wish to change the statistical interval to 1 min instead of 10 m, how can I
achieve this by modifying the source code? I have been finding the related part
of the code but I can't find it. Please help.
By the way how can I make the newly assigned executors start working more
quickly? I notice a unstable interval(sometimes long sometimes short) before my
new worker start processing data.
I notice Cluster provide these two methods for assigning executors and freeing slot.
public void freeSlot(WorkerSlot slot) public void assign(WorkerSlot slot, String
topologyId, Collection<ExecutorDetails> executors)
Say I have 20 executors in slot A, 5 exe in slot B, now I want to change one of
the executors from slot A to slot B. Do I have to shutdown both worker/slot A
and B and then restart reassigned A and B to complete my transferring just one
executor? Is there any more efficient way to do that?
If there are any violation of rules of the list please forgive my rudeness and
notify me.
thanks.
