Document basic resource requirements

[jdonenine]

Is your feature request related to a problem? Please describe.
It's not clear what system resource requirements there are to run k8ssandra.

Describe the solution you'd like
Document the resource requirements for some basic development and production scenarios.

[bradfordcp]

There are a few items to document here. This ticket describes resources for a development / production scenario, which reads as k8s cluster requirements (and potentially specs for VMs in environments where resources may be limited, eg Docker for Mac).

1. Document K8s cluster requirements - Node count, CPU, memory, and storage
2. Document K8ssandra component resource requirements / limits and how to override within values.yaml
3. What are the different values for a single worker deployment (dev environment) vs multi-worker (prod)?
4. Should we call out how binpacking multiple C* nodes per worker changes the specs?

[jsanda]

I know we expose resources for some components but not sure if we do for all. We should create additional tickets for any components for which resources are not configurable.

[jdonenine]

Yesterday when we talked about this we discussed trying to establish a resource estimate based on a few different profiles of users, @bradfordcp @jsanda do you all have any suggestions on what those profiles should be, beyond a "developer"?

We also need to figure out are we talking bare minimum resources to deploy? resources with some simulated load applied? Load applied directly? Via Stargate? There's a lot to consider here 😄

[jdonenine]

@adejanovski is going to run some performance tests with NoSQLBench to evaluate some developer focused settings.

For 1.0 our goal will be to create a measured set of resource defaults (memory, cpu, disk) that we can provide for a local developer environment and be able to document what types of tests were executed against that environment.

[adejanovski]

Based on the benchmarks and the blog post I'm writing, we have the following min requirements for Docker resources to run the whole stack (including 3 Cassandra nodes and 1 Stargate node):

• 4 cores
• 8GB RAM
• Not sure about disks, I didn't really check that part. 40GB of free space to pull the various images and put some data in seems like a decent assumption.

Reducing the number of Cassandra nodes to a single one allows to go as low as 4GB RAM instead of 8GB.

The above specs allow for a 100 ops/s workload with decent latencies.

[adejanovski]

@johnwfrancis, is that enough information for you to document this?
I can point you to the draft blog post as well for more details.

[johnwfrancis]

@adejanovski actually it's not. I've not been able to get 3 nodes running reliable with 6 cores and 8 gigs of RAM. I need the helm chart settings and any K8s config info--for instance, do you need 3 k8s workers? What should the heap/ram info be for the k8ssandra helm chart? Pretend like I'm an idiot, which, in this case, I sort of am. Once I get that, I'll run through the configuration on my machine to make sure everything is working, repeatable and reliable. Thanks!

[adejanovski]

Good point @johnwfrancis!
I need to give you the correct settings to get the stack running with 8GB RAM:

cassandra:
  datacenters:
  - name: dc1
    size: 3
  heap:
    size: 500M
    newGenSize: 250M

stargate:
  enabled: true
  replicas: 1
  heapMB: 300

With 4GB RAM, that would be:

cassandra:
  datacenters:
  - name: dc1
    size: 1
  heap:
    size: 500M
    newGenSize: 250M

stargate:
  enabled: true
  replicas: 1
  heapMB: 300

The number of workers doesn't matter performance wise, but if you have a single worker node (like with minikube), you'll need to set allowMultipleNodesPerWorker to true (defaults to false). That would then give the following values:

cassandra:
  allowMultipleNodesPerWorker: true
  datacenters:
  - name: dc1
    size: 3
  heap:
    size: 500M
    newGenSize: 250M

stargate:
  enabled: true
  replicas: 1
  heapMB: 300

[johnwfrancis]

Thanks!! So with only 4 gigs (allocated to Docker in the case of the standalone solutions like Kind/K3d/MInikube) you're limited to a single C* node, right?

[adejanovski]

Correct 👍

[johnsmartco]

@johnwfrancis completed this doc request, https://k8ssandra.io/docs/getting-started/#resource-recommendations-for-local-kubernetes-installations. Closing.