Cloud Capacity Planning

Cloud Capacity Planning

Overview

Capacity planning answers two distinct questions that are easy to conflate: "is this resource sized correctly right now" (right-sizing) and "will it still be sized correctly in N weeks given its growth trend" (forecasting). This skill covers both, across whatever cloud platform(s) an org has connected, and is deliberately conservative about calling something a risk — a capacity plan that cries wolf on every metric blip gets ignored.

This is infrastructure-substrate capacity — compute, storage, database, and cluster headroom on the platforms themselves. It is not application-level performance or SLO tracking (see devops-pack, if connected) and it is not spend (see the cloud-cost-management skill, a related but separate concern: a resource can be correctly sized and still be a cost problem, or be under-provisioned and cheap).

Discovering available tools first

Never assume which cloud platform is connected:

1. Call conduit__search_tools with a query like "list resources", "resource group", "droplet", or "quota" to discover which cloud platform connector(s) are live and their actual tool names (e.g. azure-mcp__group_resource_list, azure-mcp__quota, digitalocean__list_droplets, digitalocean__list_kubernetes_clusters, digitalocean__list_databases).
2. More than one cloud platform can be connected (an org running both Azure and DigitalOcean). Cover all connected platforms; don't stop at the first.
3. Only call concrete tools that discovery actually returned.

Key Concepts

Right-sizing signals, per platform

Platform	Over-provisioned signal	Under-provisioned signal
Azure	Resource group / subscription quota usage well below allocated quota (via azure-mcp__quota); Advisor recommendations flagging low-utilization VMs or oversized SKUs (via azure-mcp__advisor); sustained low CPU/memory/IOPS in azure-mcp__monitor metrics against an oversized SKU	Quota usage approaching the allocated limit; Advisor or azure-mcp__resourcehealth flagging throttling, sustained high utilization, or scale-limited resources
DigitalOcean	A Droplet or Database sized well above its sustained CPU/memory/disk usage; a DOKS node pool with persistently low node utilization; unattached or lightly used block storage	A Droplet or Database consistently near its CPU/memory/disk ceiling; a DOKS cluster with pods pending due to insufficient node capacity; a Database approaching connection-limit or storage-limit thresholds

Genuine capacity risk vs. normal variance

Do not flag a resource as at-risk from a single data point or a short window. Apply this discipline:

1. Require a trend, not a spike. A single hour or day of elevated utilization (batch job, deploy, traffic burst) is normal variance. A metric that has climbed over multiple consecutive observation windows (e.g., week-over-week) is a trend worth forecasting against.
2. Distinguish burst-tolerant from sustained-critical resources. A Droplet that spikes to 95% CPU for ten minutes during a nightly job is fine. A database consistently running at 85%+ storage utilization with no cleanup planned is a real risk — it degrades gracefully into an outage, not a burst.
3. State the observation window used. Always name how much history the forecast is based on (e.g., "based on the last 30 days of azure-mcp__monitor data") — a forecast built on three days of data is weaker evidence than one built on ninety, and the reader needs to know which they're getting.
4. When historical/trend data isn't exposed, say so explicitly and report current utilization as a point-in-time snapshot rather than fabricating a trend line.

Growth-trend-based forecasting

1. Pull utilization history for the resource over the longest available window the connected platform exposes.
2. Compute the trend direction and rate (e.g., "storage utilization has grown ~3%/week over the last 8 weeks").
3. Project forward to the point the resource would hit a critical threshold (e.g., 90% of allocated capacity) at the observed rate, and state that projected date as a range, not a false-precision single day — growth rates fluctuate.
4. Flag only resources whose projected exhaustion falls within a near-to-medium planning horizon (e.g., inside ~90 days) as needing near-term action; note longer horizons as "monitor, no action needed yet."

Common Workflows

Portfolio right-sizing sweep

1. Discover connected cloud platforms via conduit__search_tools.
2. Pull resource inventory (resource groups, Droplets, DOKS clusters, managed databases) per connected platform.
3. Pull utilization/quota data for each and classify: over-provisioned / right-sized / under-provisioned / insufficient data.
4. Return a ranked list — under-provisioned (real risk) first, then over-provisioned (savings/right-sizing opportunity), then a clean summary of correctly sized resources.

Capacity forecast for a resource type

1. Discover connected platforms.
2. Scope to the requested resource type (compute, storage, database, or all) per the caller's request.
3. Pull the longest available utilization history for resources of that type.
4. Apply the trend-vs-variance discipline above and produce a forecast timeline per at-risk resource, plus a "no near-term risk" summary for the rest.

Error Handling

No cloud platform connector discovered

Say so explicitly: "No cloud platform connector (Azure, DigitalOcean) is available through the gateway, so there's no capacity data to report." Do not fabricate resource data.

Platform connected but historical/trend data not exposed

Report current point-in-time utilization and state plainly that a trend-based forecast wasn't possible — do not extrapolate from a single reading.

Ambiguous resource-type scope

If asked to scope to a resource type that doesn't map cleanly onto what's connected (e.g., "database" requested but only compute platforms are connected), say so and report what is available instead of silently returning an empty result.

Best Practices

• Always discover tools before calling them — never hardcode a vendor's tool name.
• Never call a metric blip a capacity risk — require a sustained trend across multiple observation windows.
• State the observation window and data source behind every forecast.
• Separate "needs action soon" from "worth tracking, not urgent" explicitly rather than presenting one undifferentiated risk list.
• Capacity and cost are related but distinct — this skill doesn't make spend judgments; hand those to cloud-cost-management.

Related Skills

• Network Health Sweep — device/network health rather than cloud resource capacity
• Cloud Cost Management — spend anomalies and reclaimable cost; a right-sized resource can still be a cost problem and vice versa