Re: How long should metrics collection on a cluster take?

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 





On Thu, Jul 26, 2018 at 9:59 AM, Pranith Kumar Karampuri <pkarampu@xxxxxxxxxx> wrote:


On Wed, Jul 25, 2018 at 10:48 PM, John Strunk <jstrunk@xxxxxxxxxx> wrote:
I have not put together a list. Perhaps the following will help w/ the context though...

The "reconcile loop" of the operator will take the cluster CRs and reconcile them against the actual cluster config. At the 20k foot level, this amounts to something like determining there should be 8 gluster pods running, and making the appropriate changes if that doesn't match reality. In practical terms, the construction of this reconciliation loop can be thought of as a set (array) of 3-tuples: [{should_act() -> bool, can_act -> bool, action() -> ok, error}, {..., ..., ...}, ...]

Each capability of the operator would be expressed as one of these tuples.
should_act() : true if the action() should be taken
can_act() : true if the prerequisites for taking the action are met
action() : make the change. Only run if should && can.
(note that I believe should_act() and can_act() should not be separate in the implementation, for reasons I'll not go into here)

An example action might be "upgrade the container image for pod X". The associated should_act would be triggered if the "image=" of the pod doesn't match the desired "image=" in the operator CRs. The can_act evaluation would be verifying that it's ok to do this... Thinking from the top of my head:
- All volumes w/ a brick on this pod should be fully healed
- Sufficient cluster nodes should be up such that quorum is not lost when this node goes down (does this matter?)
- The proposed image is compatible with the current version of the CSI driver(s), the operator, and other gluster pods
- Probably some other stuff
The action() would update the "image=" in the Deployment to trigger the rollout

The idea is that queries would be made, both to the kube API and the gluster cluster to verify the necessary preconditions for an action prior to that action being invoked. There would obviously be commonality among the preconditions for various actions, so the results should be fetched exactly once per reconcile cycle. Also note, 1 cycle == at most 1 action() due to the action changing the state of the system.

Given that we haven't designed (or even listed) all the potential action()s, I can't give you a list of everything to query. I guarantee we'll need to know the up/down status, heal counts, and free capacity for each brick and node.

Thanks for the detailed explanation. This helps. One question though, is 5 seconds a hard limit or is there a possibility to configure it?

I put together an idea for reducing the mgmt operation latency involving mounts at https://github.com/gluster/glusterd2/issues/1069, comments welcome.
@john Still want to know if there exists  a way to find if the hard limit can be configured...
 
 

-John

On Wed, Jul 25, 2018 at 11:56 AM Pranith Kumar Karampuri <pkarampu@xxxxxxxxxx> wrote:


On Wed, Jul 25, 2018 at 8:17 PM, John Strunk <jstrunk@xxxxxxxxxx> wrote:
To add an additional data point... The operator will need to regularly reconcile the true state of the gluster cluster with the desired state stored in kubernetes. This task will be required frequently (i.e., operator-framework defaults to every 5s even if there are no config changes).
The actual amount of data we will need to query from the cluster is currently TBD and likely significantly affected by Heketi/GD1 vs. GD2 choice.

Do we have any partial list of data we will gather? Just want to understand what this might entail already...
 

-John


On Wed, Jul 25, 2018 at 5:45 AM Pranith Kumar Karampuri <pkarampu@xxxxxxxxxx> wrote:


On Tue, Jul 24, 2018 at 10:10 PM, Sankarshan Mukhopadhyay <sankarshan.mukhopadhyay@gmail.com> wrote:
On Tue, Jul 24, 2018 at 9:48 PM, Pranith Kumar Karampuri
<pkarampu@xxxxxxxxxx> wrote:
> hi,
>       Quite a few commands to monitor gluster at the moment take almost a
> second to give output.

Is this at the (most) minimum recommended cluster size?

Yes, with a single volume with 3 bricks i.e. 3 nodes in cluster.
 

> Some categories of these commands:
> 1) Any command that needs to do some sort of mount/glfs_init.
>      Examples: 1) heal info family of commands 2) statfs to find
> space-availability etc (On my laptop replica 3 volume with all local bricks,
> glfs_init takes 0.3 seconds on average)
> 2) glusterd commands that need to wait for the previous command to unlock.
> If the previous command is something related to lvm snapshot which takes
> quite a few seconds, it would be even more time consuming.
>
> Nowadays container workloads have hundreds of volumes if not thousands. If
> we want to serve any monitoring solution at this scale (I have seen
> customers use upto 600 volumes at a time, it will only get bigger) and lets
> say collecting metrics per volume takes 2 seconds per volume(Let us take the
> worst example which has all major features enabled like
> snapshot/geo-rep/quota etc etc), that will mean that it will take 20 minutes
> to collect metrics of the cluster with 600 volumes. What are the ways in
> which we can make this number more manageable? I was initially thinking may
> be it is possible to get gd2 to execute commands in parallel on different
> volumes, so potentially we could get this done in ~2 seconds. But quite a
> few of the metrics need a mount or equivalent of a mount(glfs_init) to
> collect different information like statfs, number of pending heals, quota
> usage etc. This may lead to high memory usage as the size of the mounts tend
> to be high.
>

I am not sure if starting from the "worst example" (it certainly is
not) is a good place to start from.

I didn't understand your statement. Are you saying 600 volumes is a worst example?
 
That said, for any environment
with that number of disposable volumes, what kind of metrics do
actually make any sense/impact?

Same metrics you track for long running volumes. It is just that the way the metrics
are interpreted will be different. On a long running volume, you would look at the metrics
and try to find why is the volume not giving performance as expected in the last 1 hour. Where as
in this case, you would look at metrics and find the reason why volumes that were
created and deleted in the last hour didn't give performance as expected.
 

> I wanted to seek suggestions from others on how to come to a conclusion
> about which path to take and what problems to solve.
>
> I will be happy to raise github issues based on our conclusions on this mail
> thread.
>
> --
> Pranith
>





--
sankarshan mukhopadhyay
<https://about.me/sankarshan.mukhopadhyay>
_______________________________________________
Gluster-devel mailing list
Gluster-devel@xxxxxxxxxxx
https://lists.gluster.org/mailman/listinfo/gluster-devel



--
Pranith
_______________________________________________
Gluster-devel mailing list
Gluster-devel@xxxxxxxxxxx
https://lists.gluster.org/mailman/listinfo/gluster-devel



--
Pranith



--
Pranith



--
Pranith
_______________________________________________
Gluster-devel mailing list
Gluster-devel@xxxxxxxxxxx
https://lists.gluster.org/mailman/listinfo/gluster-devel

[Index of Archives]     [Gluster Users]     [Ceph Users]     [Linux ARM Kernel]     [Linux ARM]     [Linux Omap]     [Fedora ARM]     [IETF Annouce]     [Security]     [Bugtraq]     [Linux]     [Linux OMAP]     [Linux MIPS]     [eCos]     [Asterisk Internet PBX]     [Linux API]

  Powered by Linux