[infinispan-issues] [JBoss JIRA] (ISPN-6879) Calculate (and expose) minimum number of nodes for data in Infinispan

[Sebastian Łaskawiec (JIRA)]

    [ https://issues.jboss.org/browse/ISPN-6879?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13481231#comment-13481231 ] 

Sebastian Łaskawiec commented on ISPN-6879:
-------------------------------------------

{quote}
I am not certain of the benefit in this case. Even if you go from 100 nodes down to 10, you would have lost ~90% of your entries and there is no way to know which ones were lost and which ones were kept.
{quote}

hmmm that's actually an interesting use case. To be honest I thought that if you scale gradually from 100 to 10 (killing one node at a time and waiting until rebalance happens) you will preserve the data. I had this impression since Kubernetes generates SIGINT when killing a node and waits ~30 second till it shuts down. I noticed that we trigger rebalance on each SIGINT (since Wildfly catches SIGINT and shuts all services down). 

So I probably was wrong ([~dan.berindei] could you please confirm this?) or the use case with number of owner = 1 doesn't make sense for this feature. We should always aim for number of owners >= 2.

{quote}
As long as numOwners < numNodes, I don't see the real benefit as you would have to throttle the nodes going down to guarantee that data isn't lost. I personally don't see a user wanting to sit there for minutes just to shut down a subset of nodes. More than likely they would want to say I want X nodes. Can we not (OpenShift or us) shut them down in an orderly fashion to do this instead? This seems much safer and wouldn't have as many user input errors.
{quote}

This one is actually easy - this is the default behavior of {{StatefulSets}}. The controller kills one node at a time and waits till cluster stabilizes.

Let me add one more aspect here - scaling based on custom metrics will *probably* need to figure out the lowest number of nodes to operate with given dataset. Without it, we might see some weird behavior. So maybe there are better ways to tell it - "Here is our minimum number of nodes. Never, ever go below that."

> Calculate (and expose) minimum number of nodes for data in Infinispan
> ---------------------------------------------------------------------
>
>                 Key: ISPN-6879
>                 URL: https://issues.jboss.org/browse/ISPN-6879
>             Project: Infinispan
>          Issue Type: Feature Request
>          Components: Cloud Integrations, Server
>            Reporter: Sebastian Łaskawiec
>            Assignee: William Burns
>
> With Kubernetes autoscaling we need to be able to tell what is the minimum amount of nodes necessary for hosting data (probably some sort of size + number of nodes estimation).



--
This message was sent by Atlassian JIRA
(v7.5.0#75005)