infra-scaling

Read the current tfvars first, then gather 7-day Datadog evidence on CPU, request counts, latency, and errors. If CPU is low and health is green, raise the RPM target or narrow the boundary; if the service sits at its minimum, lower the minimum task count instead.

Check whether services are floor-bound at their minimum task count, since raising RPM targets saves nothing there. Lower accidental idle floors after verifying HA and compliance requirements, and raise the max-capacity multiplier to preserve burst ceiling.

Not automatically. Inspect full APM traces first; if slow spans are dominated by ClickHouse queries, lowering RPM targets will not reduce query cost or duration. Only reduce RPM when evidence shows container-local saturation or event-loop delay.

Scale-up threshold is rpm_target plus ceil(rpm_target times boundary), and scale-down is rpm_target minus the same. Compute markers with tofu console rather than mental math, because Terraform float behavior can cause off-by-one errors.

Prod web-ingestion keeps a minimum of 3 containers, and 3 is also an acceptable prod minimum for web and web-iso when no higher HA, isolation, or compliance floor is required. The web_ingestion_instances variable acts as the autoscaler minimum.

Only after presenting a human-review table and receiving explicit approval of row IDs and actions. It creates one ticket per scaling adjustment, never assigns tickets without an explicit assignee, and never creates tickets for Valeriy.