137 lines
4.8 KiB
Markdown
137 lines
4.8 KiB
Markdown
**Drop 2 – Store‑in‑a‑Box: Hardware & Proxmox in Plain English**
|
||
*Part of the LinkedIn series “Edge Renaissance: putting compute—and the customer—back where they belong.”*
|
||
|
||
---
|
||
|
||
### ☕ Executive espresso (60‑second read)
|
||
|
||
* **Three shoebox PCs ≈ one mini‑cloud.** For <\$6k/site you get HA, live‑migration, snapshots—no VMware tax.
|
||
* **It’s not about servers for servers’ sake.** This kit exists to shave 100 ms off every click and keep kiosks alive when the WAN dies.
|
||
* **Plain English stack:** Proxmox = the “operating system for your private cloud.” KVM runs full VMs, LXC runs lightweight containers, Ceph keeps copies of your data on all three boxes.
|
||
|
||
> **Bottom line:** You already power closets in every store. Drop in three nodes, wire them once, and you’ve got the platform to out‑Amazon Amazon on customer obsession—without their cap‑ex.
|
||
|
||
---
|
||
|
||
## 1️⃣ What actually goes in the closet?
|
||
|
||
```
|
||
[ Node A ] [ Node B ] [ Node C ]
|
||
├─ CPU: 8–16 cores (Ryzen / Xeon-D)
|
||
├─ RAM: 64–128 GB
|
||
├─ NVMe: 2 × 1–2 TB (mirrored)
|
||
└─ NIC: 2 × 10/25 GbE
|
||
|
||
[ Switch ]
|
||
├─ 10/25 GbE for cluster replication
|
||
└─ 1 GbE uplink to store LAN/WAN
|
||
|
||
[ UPS ] ≈ 1500 VA line‑interactive unit
|
||
```
|
||
|
||
Space: half a rack or a wall‑mount cabinet. Power: <500 W total under load.
|
||
|
||
---
|
||
|
||
## 2️⃣ Bill of materials (copy‑paste ready for LinkedIn)
|
||
|
||
```
|
||
GOOD (≈ $3.5k)
|
||
• 3 × Mini‑PC (Ryzen 7 / 64 GB / 2 × 1 TB NVMe) … $900 ea
|
||
• 1 × Fanless 10 GbE switch (8‑port) … $400
|
||
• 1 × 1500 VA UPS … $300
|
||
|
||
BETTER (≈ $5.5k)
|
||
• 3 × SFF server (Xeon‑D / 96 GB / 2 × 2 TB NVMe) … $1,400 ea
|
||
• 1 × 12‑port 25 GbE switch … $700
|
||
• 1 × Smart PDU + 2U wall rack … $300
|
||
|
||
BEST (≈ $8k+)
|
||
• 3 × Edge GPU nodes (RTX A2000 / 128 GB RAM) … $2,200 ea
|
||
• 1 × 25 GbE switch + SPF28 optics … $900
|
||
• Redundant UPS + environmental sensors … $500
|
||
```
|
||
|
||
*(Swap SKUs as vendors change—targets are core counts, RAM, NVMe, and dual NICs.)*
|
||
|
||
---
|
||
|
||
## 3️⃣ Proxmox, demystified
|
||
|
||
* **Proxmox VE (Virtual Environment):** The web UI + API that manages everything. Think “VMware vSphere, but open‑source.”
|
||
* **KVM VMs:** Full OS instances (Windows POS, legacy apps).
|
||
* **LXC containers:** Lightweight Linux “jails” for APIs, caches, edge functions.
|
||
* **Ceph storage:** Each disk contributes to a shared pool; lose a node, data’s still there.
|
||
* **Proxmox Backup Server (PBS):** Built‑in, deduped backups to another box or S3 bucket.
|
||
|
||
> Translation: High availability and snapshots without buying a hyper‑converged appliance.
|
||
|
||
---
|
||
|
||
## 4️⃣ How resilience actually works
|
||
|
||
```
|
||
Normal: All 3 nodes active → Ceph keeps 3 copies of data
|
||
Failure: Node B dies → workloads live‑migrate to A & C
|
||
Network: WAN drops → local DNS/cache/APIs keep serving
|
||
Recovery: Replace/repair node → Ceph heals automatically
|
||
```
|
||
|
||
No one calls IT; the store keeps ringing sales, kiosks keep scanning, mobile app keeps answering.
|
||
|
||
---
|
||
|
||
## 5️⃣ Install & bootstrap in five steps
|
||
|
||
```bash
|
||
# 1. Image USB with Proxmox VE ISO and install on each node
|
||
# 2. Create a cluster on the first node
|
||
pvecm create store-$SITE_ID
|
||
|
||
# 3. Join the other nodes
|
||
pvecm add <IP_of_first_node>
|
||
|
||
# 4. Configure Ceph (3 mons, 3 OSDs)
|
||
pveceph install
|
||
pveceph createmon
|
||
pveceph osd create /dev/nvme1n1
|
||
|
||
# 5. Push your golden VMs/containers via Ansible/Terraform
|
||
ansible-playbook edge_bootstrap.yml -e site=$SITE_ID
|
||
```
|
||
|
||
*(We’ll publish the full playbook in Drop 6.)*
|
||
|
||
---
|
||
|
||
## 6️⃣ “But do we really need three boxes?”
|
||
|
||
* **2 nodes** = cheaper, but no true quorum. You’ll need an external witness (tiny VPS).
|
||
* **3 nodes** = true HA + Ceph replication. This is the sweet spot.
|
||
* **1 node** = pilot only (no HA, but fine for a proof‑of‑value store).
|
||
|
||
---
|
||
|
||
## 7️⃣ Tie it back to customer obsession (not just cost)
|
||
|
||
* **Faster everything:** APIs, PDP images, kiosk menus—served from 50 feet away.
|
||
* **Always on:** WAN outage? Your store experience doesn’t blink.
|
||
* **Personal, local, real:** The same cluster that runs inventory logic personalises promos on the PDP—because it has the freshest stock data.
|
||
|
||
---
|
||
|
||
### ✅ This week’s action list
|
||
|
||
1. **Pick your tier (Good/Better/Best)** and price it for 5 pilot stores.
|
||
2. **Order one cluster** and set it up in a lab/back office.
|
||
3. **Move 2 workloads first:** image cache + `/inventory` API. Measure the latency drop.
|
||
4. **Write a one‑pager** for execs: “Cost of three nodes vs. cost of 100 ms latency.”
|
||
|
||
---
|
||
|
||
### Next up ➡️ **Drop 3 – DIY CDN: Serving shoppers from 50 feet away**
|
||
|
||
We’ll turn this cluster into a location‑aware CDN so your digital customers get the same sub‑30 ms treatment.
|
||
|
||
*Stay subscribed—your broom closets are about to earn their keep.*
|