Lumen Analytics: A 40 Percent Overrun and No One to Call
Lumen Analytics runs a steady baseline of 200 m6i instances for ETL that must run 24/7, plus a nightly rendering job of 600 instances that can tolerate interruption and restarts. Finance approved a 12-month cloud budget. Last month's bill was 40% over forecast. The cause was not compute: a NAT Gateway processed 90 TB after a misconfigured job pulled data from S3 through the NAT, and 25 TB crossed Availability Zones between the web tier and database. The company is on Basic Support and experienced a six-hour production outage last quarter with no one to call. The CFO wants commitments only where risk is low, and the CTO wants a support plan that covers production impact.
1. Design the purchasing strategy for the baseline ETL fleet and the nightly rendering job. Explain the trade-offs.
Read the worked answer
Baseline ETL (200 m6i, 24/7, predictable): use a one-year Compute Savings Plan with no upfront payment. It discounts EC2, Fargate, and Lambda usage and remains flexible across instance family, size, and Region, which suits a CFO who is cautious about long commitments. If the instance family is genuinely fixed for years, an EC2 Instance Savings Plan gives a deeper discount but trades away that flexibility. A three-year term would save more but conflicts with the stated risk preference.
Nightly rendering (600 instances, interruption-tolerant): use Spot capacity with a diversified set of instance types and Availability Zones, a capacity-optimized allocation strategy, and checkpointing so an interruption restarts rather than loses work. Keep a small On-Demand buffer for jobs that must finish. Do not put baseline ETL on Spot; a two-minute interruption notice is incompatible with 24/7 processing.
Mistake to avoid: committing to cover the nightly peak. Commitments should cover the steady floor, not the burst. Reserved Instances still make sense for steady database workloads, but not for render nodes.
2. Diagnose the network charges and propose changes that reduce them without breaking availability.
Read the worked answer
NAT Gateway charges have two parts: an hourly charge and a per-GB data processing charge. Pulling S3 data through a NAT Gateway pays that processing charge unnecessarily. Create an S3 gateway VPC endpoint, which is free, and route S3 traffic through it. Add interface endpoints for other AWS services the workloads use frequently. Fix the misconfigured job so it reads from the endpoint path or from within the same Region.
Cross-AZ charges: data transferred between Availability Zones is billed in each direction. The 25 TB between web tier and database reflects a chatty design. Options: keep tightly coupled components in the same AZ where the availability trade-off is acceptable, reduce chatty calls through caching, and use same-AZ read replicas for read traffic. Do not eliminate multi-AZ entirely; failover capability is worth some transfer cost. Instead, remove unnecessary cross-AZ chatter while preserving a standby.
Governance: enable cost allocation tags, AWS Budgets with alerts, and Cost Anomaly Detection. A 90 TB surprise should have triggered an alert on day one, not at month end. Review VPC Flow Logs for top talkers.
3. Recommend a support plan and justify it against the alternatives.
Read the worked answer
Recommend Business Support at minimum. It provides 24/7 access to phone, chat, and email, a response target of under one hour for a production system down, and the full set of Trusted Advisor checks, including security and cost optimization. The six-hour outage with no one to call is precisely the failure mode Business Support is designed to prevent.
Alternatives: Developer Support is business-hours email only and is unsuitable for a production revenue system. Enterprise On-Ramp adds a 30-minute response for business-critical outages and access to a pool of Technical Account Managers, which is reasonable if Lumen wants architectural guidance and faster escalation. Enterprise Support adds a dedicated TAM and a 15-minute business-critical response, but the cost is hard to justify for a company still stabilizing its bill.
Decision rule: production impact and revenue loss set the floor. If a six-hour outage is plausible again, the cost of Business or On-Ramp is small relative to the loss. The CTO should also note that support plan charges are a predictable operating cost and should be included in the budget rather than treated as overhead discovered later.