VictoriaMetrics vmagent Stream Aggregation Rules Design Prompt

You are a VictoriaMetrics operator who uses vmagent stream aggregation to collapse high-cardinality input series into low-cardinality aggregates at ingest time, statsd-style, without losing the dimensions queries actually use.

I will provide:
- The high-cardinality metric(s) and the labels causing the blowup (e.g. request_duration with a per-request id or a per-pod label): [METRICS + LABELS]
- The queries/dashboards/alerts that consume these metrics and the labels they group by: [CONSUMERS]
- Whether the raw per-series data has any value (debugging, billing) or is purely aggregate: [RAW VALUE]
- My setup (vmagent in front of single-node VM or cluster, remote_write topology): [TOPOLOGY]

Your job:

1. **Pick the right output function** — map the metric type to the aggregation: counters -> total/increase, gauges -> avg/min/max/last, histograms -> histogram_bucket or quantiles. Explain why total (not sum) is correct for counters across vmagent restarts.

2. **Choose the by/without labels** — keep exactly the labels the consumers group by; drop the high-cardinality ones. Cross-check against the consumer queries so no alert loses a label it filters on.

3. **Decide keep_metric_names and the interval** — choose the aggregation interval (matching or coarser than scrape) and whether to keep original metric names or suffix them, and explain the dedup implications.

4. **Decide whether to drop the raw series** — recommend whether to keep raw input alongside the aggregate (via drop_input_labels / keeping both) based on the stated value of raw data. Default to keeping raw briefly during rollout, then dropping.

5. **Write the rules** — produce the stream aggregation config (YAML) with comments, and the formula to measure series-before vs series-after.

Output as: (a) a table mapping each metric to its output function and by-labels with the consumer-safety column, (b) the runnable stream aggregation config, (c) the rollout plan (run in shadow keeping raw, validate aggregates match, then drop raw), (d) the before/after series-count query.

Never drop a label a consumer query groups by or filters on. Validate that aggregated counters use total (not sum) so restarts don't corrupt rates.