Cloud Operations, Monitoring, Automation and Cost Optimization
What to know
Operations questions test whether you can see, react, and pay less.
See: CES collects metrics (basic free metrics at coarse granularity; detailed metrics at 1-minute granularity, billed) and needs the agent installed for guest-OS metrics such as memory and disk usage. LTS centralises log collection and query/analysis, with transfer to OBS for long retention. AOM/APM cover application-level operations and distributed tracing. CTS records management-plane API calls for audit and forensics.
React: alarms in CES trigger SMN notifications or direct actions such as scaling. Auto Scaling policies (scheduled, periodic, alarm-based) provide the elastic loop; FunctionGraph provides event-driven automation for glue tasks like snapshot creation, tag enforcement, or scheduled remediation. An alarm without an action, an owner, or a runbook is monitoring theatre.
Pay less: choose the right purchase model (yearly/monthly subscription for predictable baselines, pay-per-use for spiky load, spot for interruptible batch); right-size over-provisioned flavors based on observed CES metrics; schedule shutdown of non-production environments outside working hours; apply OBS lifecycle rules to move cold data to archive tiers; delete orphaned EVS volumes, unattached EIPs, and idle load balancers.
Traps
Stopping an ECS does not stop billing for its EVS disks or EIP. A pay-per-use EVS volume continues to bill while detached. Detailed monitoring is a billed feature, and without the agent you simply will not see memory utilisation - a common reason right-sizing exercises fail. Cost allocation requires consistent tagging; without tags you cannot do chargeback.
Rule of thumb
Run a monthly hygiene review with four checks: unattached volumes, idle EIPs, instances below ~10% CPU, and objects still in Standard that have not been read in 90 days. These four recover most avoidable spend.