Hi Matthieu,
thanks for you answers, they helped me a lot.
Regards,
Benjamin
On 01/03/2013 04:16 PM, Matthieu Morel wrote:
Hi Benjamin, some answers inline:
On Jan 3, 2013, at 15:55 , Benjamin Süß wrote:
in the overview about s4 v0.5 I there is a point mentioning "dynamic
application deployment". Unfortunately I could not find any further
information about this so I have some questions:
1.) In general, does this mean new applications can be added at
cluster runtime?
We cannot have multiple S4 applications on the same logical cluster.
But we can start with "empty" nodes and dynamically deploy an
application. This model facilitates porting the platform to other
infrastructures (e.g. Yarn), while keeping a common deployment scheme .
2.) If so, how about updating existing applications
Currently we have to restart the nodes, and update the app reference
in Zookeeper. (The existing classloading machinery (inspired from
Hadoop's) could probably be extended to support automatic updating
though).
3.) If it is possible to update existing applications, what about
data loss during updates?
You can recover previous state through checkpointing and recovery, but
pending messages are lost.
Hope this helps,
Matthieu
