CloudPulse

Architecture

Services emit RED metrics and resource samples. A collector aggregates time series, joins a dependency map, and flags anomalies against simple thresholds. Cost is rolled up by service from illustrative unit rates. This build uses a synthetic, seeded dataset—no cloud account is connected.

Instrumentation → metric aggregation → service health model → charts and dependency map → anomaly rules → incident evidence pack.

Decisions

• — One synthetic dataset feeds charts, totals, alerts, and the incident view so widgets cannot contradict each other.
• — Time-range and service filters are query parameters of the same series, not separate mock screens.
• — Cost is labeled as modeled example spend derived from request volume and a unit rate.

Trade-offs

• — Simple threshold alerts are explainable and they miss slow-burn saturation that needs baselining.
• — Service-level rollups are scannable and they hide noisy-neighbor effects inside a shared cluster.
• — Modeled cost from RPS is directionally useful in a demo and is not a bill.

Failure modes

• — A downstream dependency can raise p95 while the selected service's error rate stays flat.
• — Sparse windows look healthy because there is not enough traffic to trip a rate alert.
• — Cost anomalies without request context invite the wrong remediation (scale vs. leak vs. retry storm).

Evaluation

Totals, error rates, and cost figures are computed from the seeded series for the selected filters. They are demo calculations, not live infrastructure.

Production extensions

Proposed — not implemented in this portfolio.

• — Ingest Prometheus and OpenTelemetry and keep the same investigation UI.
• — Join traces to metric spikes for a true evidence pack.
• — Replace modeled unit cost with CUR / actual billing exports.