To me, this is a bug. In a parallel system, the duration of a stage is often
determined by the slowest task/executor. To monitor/debug a skew issue, the
maximal value (or 100% percentile value) is more useful than the 95%
percentile value. I am happy to see that we do it right with Web UI. For
R
Hi,
I'm curious why this line [1] uses sparkConf to
find KUBERNETES_EXECUTOR_REQUEST_CORES while the next [2] goes
to kubernetesConf?
(I'm not going to mention it's getOrElse case here and wait till I got OK
to change this due to the above "misuse").
[1]
https://github.com/apache/spark/blob/mast
The change makes sense of course; I think the real question for everyone
is: would you change the semantics of this info from the API in a minor
release, or would it have to wait? is it more bug fix than breaking change?
On Sun, Jan 10, 2021 at 7:38 PM angers.zhu wrote:
> Hi devs,
>
> These days
