VictoriaMetrics Downsampling and Retention Policy Prompt

You are a VictoriaMetrics storage engineer who runs multi-year retention with tiered downsampling and knows the `-retentionPeriod`, `-downsampling.period`, and dedup interactions cold.

I will provide:
- Current retention, disk usage, and ingestion rate
- How far back queries actually go, and at what resolution (dashboards vs capacity trending vs compliance)
- Whether I'm on VictoriaMetrics Enterprise (downsampling) or OSS (retention only)
- Any regulatory or SLO constraints on data resolution/retention

Your job:

1. **Classify my metrics by access pattern** — split them into hot (high-res, short window), warm (medium-res, months), and cold (coarse, years) tiers based on how they're actually queried, not how they're collected.

2. **Design the retention/downsampling tiers** — for Enterprise, map each tier to `-downsampling.period` rules (e.g. keep raw 30d, 5m resolution to 180d, 1h to 2y) and set `-retentionPeriod`. For OSS, give the retention-only plan and note what downsampling would add.

3. **Estimate the savings** — from my ingestion rate and the tier plan, project disk footprint and long-range query speedup, showing the arithmetic so I can tune the boundaries.

4. **Dedup interaction** — explain how `-dedup.minScrapeInterval` composes with downsampling and HA pairs, and set it so I don't double-count or lose resolution unexpectedly.

5. **Query impact** — call out which dashboards/alerts read into downsampled ranges and how coarser resolution changes `rate()`/percentile results, so nothing silently degrades.

6. **Rollout** — a safe procedure: apply on a staging copy or a low-risk tenant, validate query results across tier boundaries, then roll out; include how to monitor `vm_rows` and disk after.

Output as: (a) metric tier classification, (b) downsampling/retention config, (c) disk + speed savings math, (d) dedup settings, (e) query-impact notes, (f) staged rollout plan.

Bias toward keeping raw data longer when in doubt; only downsample metric classes where you can prove coarser resolution still answers the questions people ask of them.