[CentOS-devel] What should I do in my cluster? CentOS streams dilemma.

Fri Dec 25 23:04:46 UTC 2020
redbaronbrowser <redbaronbrowser at protonmail.com>

On Friday, December 25, 2020 2:02 PM, Gordon Messmer <gordon.messmer at gmail.com> wrote:

> On 12/25/20 5:24 AM, redbaronbrowser via CentOS-devel wrote:
> > Con of using CentOS Stream is it .. packages will be released before getting the same level of QA of RHEL 8 or CentOS 8.
> Several Red Hat and CentOS engineers have said that each CentOS Stream
> package will have passed QA before they are published.

How invested are they on that?

How often will a package be released on RHEL at the same exact time it is published on Stream?

We played this game in the pre-FESCo days of Fedora as well.

If the gamble works out then it is because they have been doing everything great.

For the packages the gamble does not work out on then it is proof people should have bought RHEL.

It is a win-win for RH/Stream engineers to make claims because they are gambling with other people's chips instead of their own.

CentOS engineers also said there was a firewall between CentOS and RHEL but we still got Brian "Bex" Exelbierd on the governing board.  It turns out they say a lot of things and not all of them are true.

My opinion still stands, not all releases to Stream will be moved as-is into RHEL.  At some point there will be a release to Stream in which the package will be updated again before RHEL is willing to release it.

I would say the chances of Stream ending up on a Top 500 cluster is about the same as Windows Insider Slow Ring builds.  The methodology seems to be similar.

However, RHEL 8 is a great product.  Gluster is also a great product but has a slight learning curve and benefits from Red Hat support.

If your application resource usage is mostly CPU bound with some amount of disk reads and low amount of disk writes, then a gluster may provide the required performance.  Combine that with a pair of 10Gbps switches providing bonded/LACP to each of the nodes, the latency/bandwidth should be minimal.

My main point is the question of finalizing what needs to be "installed" on the cluster shouldn't matter for the majority of node.

Nether Fedora or Stream are what I would recommend for the head nodes.  I would recommend making sure to get support for those.

The other nodes should be easy to swap what is booted over the network or installed via anconda kickstart automation.

If there is an application that would benefit from Stream on the nodes then put that on the nodes for that application run.  If there is an application that would benefit from Fedora on the ndoes then put that on the nodes for the run of that application.  If there is an application that would benefit from Debian on the nodes, then put that on during that run.

If you are making it on the bottom of the Top 500 list using something like System76 1U AMD servers with 40 cores per server (full disclosure: I personally have not used System76 servers), then you are looking at 14 cabinets full servers (if not more).  What is installed on the majority of those nodes should be automated and reinstalls should be "cheap."  The question of choosing what to install for the long term sounds a little odd.