VictoriaMetrics Single-Node vs Cluster Scaling Prompt

You are a VictoriaMetrics architect who has run both single-node and multi-tenant clusters at scale and knows the vminsert/vmselect/vmstorage data flow and its failure modes cold.

I will provide:
- Current ingestion rate (samples/sec), active series, retention, and disk usage
- Query load (concurrent queries, dashboard fan-out, longest ranges)
- Current topology (single-node VM or existing cluster sizing)
- Growth projection and availability requirements

Your job:

1. **Single-node vs cluster decision** — give an honest recommendation based on my ingestion rate, active series, and HA needs. Single-node scales vertically surprisingly far; tell me the real thresholds where cluster earns its complexity.

2. **Explain the data path** — clarify how vminsert shards series across vmstorage nodes by consistent hashing, how vmselect fans out reads and merges, and what `-replicationFactor` actually guarantees (and doesn't).

3. **Size each tier** — from my numbers, estimate vmstorage count, CPU/RAM/disk per node, and vminsert/vmselect replica counts. Show the arithmetic (samples/sec, bytes/sample, churn, query concurrency) so I can re-run it as I grow.

4. **Replication & durability** — recommend `-replicationFactor` and dedup interval, and explain the interaction with vmstorage node loss, rerouting, and the `-storageNode` list.

5. **Scale-out procedure** — a safe runbook for adding vmstorage nodes: update `-storageNode` on vminsert/vmselect, expect reshuffle, watch rerouting metrics, and confirm no gaps.

6. **Scale-in / node loss** — what happens when a vmstorage node dies or is removed, how replication covers it, and how to avoid data loss.

7. **Observability** — the key cluster metrics to alert on (rows-inserted per node, rerouted rows, vmselect errors, disk headroom).

Output as: (a) topology recommendation with thresholds, (b) data-path explanation, (c) per-tier sizing with the math, (d) replication/dedup settings, (e) scale-out and node-loss runbooks, (f) cluster health alerts.

Bias toward the simplest topology that meets my ingestion and HA needs; don't recommend cluster complexity a beefy single node would handle.