Pi Cluster

[Temmpp]

Hi,

 

So I am creating my first home lab project and after researching I've decided to start using a raspberry pi 5. I've also 

 

I am going to be running 3 raspberry pi's.

1 - Pi-Master (8gb ram): Will be running CouchDB for obsidian, K3s and Pi hole. This will have a Pi Hat (SSD)

2 - Pi-Ai (16gb ram): will be running an ollama model, n8n, Qdrant, Open WebUI. This will have a Pi Hat (SSD)

3 - Pi-worker: will be running Uptime Kuma, and have space for me to experiment with. 

 

Each feature will be in its own docker container at first, then I will implement K3s when I 

 

Will be using tailscale for external connection to it.

 

This is my set up so far, I'm yet to actually assemble everything, but items have been ordered.

 

My question is, what should I experiment with on Pi-worker? Any projects I should do? Any improvements or suggestions?

 

Thanks!

[Levent]

Waste of money, why do you even need a "cluster" to begin with.. Get a ex office machine, slap 32G/64G ram in it. Install proxmox and do everything in there. 

[igormp]

1 hour ago, Temmpp said:

Hi,

 

So I am creating my first home lab project and after researching I've decided to start using a raspberry pi 5. I've also 

 

I am going to be running 3 raspberry pi's.

1 - Pi-Master (8gb ram): Will be running CouchDB for obsidian, K3s and Pi hole. This will have a Pi Hat (SSD)

2 - Pi-Ai (16gb ram): will be running an ollama model, n8n, Qdrant, Open WebUI. This will have a Pi Hat (SSD)

3 - Pi-worker: will be running Uptime Kuma, and have space for me to experiment with. 

 

Each feature will be in its own docker container at first, then I will implement K3s when I 

 

Will be using tailscale for external connection to it.

 

This is my set up so far, I'm yet to actually assemble everything, but items have been ordered.

 

My question is, what should I experiment with on Pi-worker? Any projects I should do? Any improvements or suggestions?

 

Thanks!

Why a "master" since all of those are independent? I guess this terminology will make more sense once you setup K3s across all devices, with the master responsible for the control place.

 

Otherwise, why not move n8n, qdrant and openwebui to the pi-worker, leaving the Pi-AI responsible solely with the ollama model? You could also do llama.cpp instead of ollama, but that's just my personal preference.

[Temmpp]

3 minutes ago, igormp said:

Why a "master" since all of those are independent? I guess this terminology will make more sense once you setup K3s across all devices, with the master responsible for the control place.

Yeah that's the idea. I know they're all technically independent, but since the master pi will be running K3s and managing that, I thought it would make sense to call it that. 
 

6 minutes ago, igormp said:

Otherwise, why not move n8n, qdrant and openwebui to the pi-worker, leaving the Pi-AI responsible solely with the ollama model? You could also do llama.cpp instead of ollama, but that's just my personal preference.

My idea behind keeping n8n on the pi-ai was to keep all AI things on one machine, maybe if I decided to expand it in future I can move over that stuff to another worker node.

 

11 minutes ago, Levent said:

Waste of money, why do you even need a "cluster" to begin with.. Get a ex office machine, slap 32G/64G ram in it. Install proxmox and do everything in there. 

Yeah that's a fair point, if I only cared about raw power/value then yeah an ex-office machine would probably make more sense. But this is my first home lab project, and the main reason I’m doing it this way is to actually learn how to manage multiple nodes, split services up properly, use Docker, and get into K3s. A single machine would be good for running everything and cheaper, but it wouldn’t really teach me the same stuff. So for me this project is less about getting the most performance for the money, and more about learning and building something properly.

[Levent]

13 minutes ago, Temmpp said:

Yeah that's a fair point, if I only cared about raw power/value then yeah an ex-office machine would probably make more sense. But this is my first home lab project, and the main reason I’m doing it this way is to actually learn how to manage multiple nodes, split services up properly, use Docker, and get into K3s. A single machine would be good for running everything and cheaper, but it wouldn’t really teach me the same stuff. So for me this project is less about getting the most performance for the money, and more about learning and building something properly.

Hence why I said proxmox, you can virtualize everything. You dont have to waste money. There is a smarter way of doing it and there is this.

[Temmpp]

13 minutes ago, Levent said:

Hence why I said proxmox, you can virtualize everything. You dont have to waste money. There is a smarter way of doing it and there is this.

Yeah, I get your point, and I agree a single machine with Proxmox would be the better option if my goal was just cost and efficiency.

For me specifically though, the point of this build is learning things a single box doesn’t teach in the same way, managing separate physical nodes, real network communication between machines, service placement, node failure/recovery, and later K3s across actual hardware. So while it may not be the cheapest route, it fits what I’m trying to learn much better.

[Kilrah]

Maybe look into multiple of them old business machines, they'll likely be cheaper and way less limiting.

LLM on a Pi5 would be atrocious...

 

[Temmpp]

9 minutes ago, Kilrah said:

Maybe look into multiple of them old business machines, they'll likely be cheaper and way less limiting.

LLM on a Pi5 would be atrocious...

 

Yeah true, its not going to be fast by all means, this is just learning so when I do it on a better machine later on I understand I have a better understanding and able to plan better. I have an old gaming PC (GPU 4060 8GB) that I lent to my friend, and when I get that back in the next couple weeks I'll wipe that and run a local LLM on there, it wont be the best but it'll be a significant improvement.

[igormp]

2 hours ago, Temmpp said:

My idea behind keeping n8n on the pi-ai was to keep all AI things on one machine, maybe if I decided to expand it in future I can move over that stuff to another worker node.

"all things AI" is kinda hard to consolidate into a single thing because you'll be sharing CPU and RAM for those services.

Things like OpenWebUI are not CPU/RAM intensive at all and just connect to your OpenAI API endpoint, so it can be run from any device. Qdrant requires memory disk storage for all of your embeddings, and may even eat up quite some CPU depending on your indexing configs. As an aside, I personally dislike Qdrant and would go with some other option, like pgvector (or any other PG vector add-on) or Chroma.

N8n really depends on what kind of nodes/workflows you'll be running. I can use no resources to requiring multiple nodes.

Running LLMs might eat up an entire Pi by itself depending on what kind of model you'll be running, thus it'd be better to dedicate a node solely for this purpose.

 

47 minutes ago, Kilrah said:

Maybe look into multiple of them old business machines, they'll likely be cheaper and way less limiting.

LLM on a Pi5 would be atrocious...

There are many smaller LLMs that are somewhat good and can be run on a Pi without being too painful. For learning purposes that's a nice experiment.

[Temmpp]

13 minutes ago, igormp said:

Running LLMs might eat up an entire Pi by itself depending on what kind of model you'll be running, thus it'd be better to dedicate a node solely for this purpose.

Yeah true, I mean this is my first time doing anything like this... so its all completely new to me, I'm using this as a chance to build up my skills to help further my career.

 

The plan for this project is to learn and experiment on the Pi Cluster, figure out what best uses I can get from it, figure out limitations, learn about Kubernetes (K3c in my case), docker and a bit about networking. And then when I eventually get my old PC back, I can then set that up as a server and put the more intensive processes on there (So for example, K3c into Kubernetes, running a better more powerful LLM then dedicating the old pi-ai to n8n for instance).

[igormp]

36 minutes ago, Temmpp said:

Yeah true, I mean this is my first time doing anything like this... so its all completely new to me, I'm using this as a chance to build up my skills to help further my career.

 

The plan for this project is to learn and experiment on the Pi Cluster, figure out what best uses I can get from it, figure out limitations, learn about Kubernetes (K3c in my case), docker and a bit about networking. And then when I eventually get my old PC back, I can then set that up as a server and put the more intensive processes on there (So for example, K3c into Kubernetes, running a better more powerful LLM then dedicating the old pi-ai to n8n for instance).

If the objective is learning, then get your hands dirty ASAP and don't bother asking too many questions, just implement things, see what works good or not and try to keep improving.

[x3ky8]

I know I'm a little new here and you asked a different question. Just thought i would throw out there with What you are wanting to do i think you could be better served using something like a Zima Board. Its in the $200 USD range but with the rising cost of all hardware (with some models of the PIs in that range) I think the zima board is better bang for your buck for example

- x86 Architecture
- Higher Single Threaded performance
- Two native SATA ports 
- eMMC or M.2 storage 
- Expandable for extra GPU 

I also believe that you can use proxmox. 

Anyway hope this helps! Good luck on your homelab adventure! 

[Nikolithebear]

On 3/30/2026 at 6:34 AM, Temmpp said:

Hi,

 

So I am creating my first home lab project and after researching I've decided to start using a raspberry pi 5. I've also 

 

I am going to be running 3 raspberry pi's.

1 - Pi-Master (8gb ram): Will be running CouchDB for obsidian, K3s and Pi hole. This will have a Pi Hat (SSD)

2 - Pi-Ai (16gb ram): will be running an ollama model, n8n, Qdrant, Open WebUI. This will have a Pi Hat (SSD)

3 - Pi-worker: will be running Uptime Kuma, and have space for me to experiment with. 

 

Each feature will be in its own docker container at first, then I will implement K3s when I 

 

Will be using tailscale for external connection to it.

 

This is my set up so far, I'm yet to actually assemble everything, but items have been ordered.

 

My question is, what should I experiment with on Pi-worker? Any projects I should do? Any improvements or suggestions?

 

Thanks!

 

 

 

PI's are a good idea if you already have them.  I won't try to force the idea of other options on you, but office PCs may be more economic here.

Grabbing 3 mini desktops like HP Elitedesk Mini's would be around 50-80 watts for the cluster of them.  From there you'd get more horse power, more control, and give you more flexibility.

But, that being said if you already have the Pi's and just want to tinker more power to you.