PATTERN Cited by 1 source

Two-axis horizontal + vertical autoscaling

Pattern

Combine request-based horizontal scaling (add/remove replicas instantly as traffic shifts) with resource-based vertical scaling (adjust per-replica concurrency as the workload's resource profile becomes clear) in a single controller where the vertical axis feeds the horizontal formula.

desired_replicas = ceil(active_requests / target_concurrency)
                                            ↑
                                     set by vertical axis

The horizontal axis handles availability and latency (the fast decision). The vertical axis handles efficiency (the steady-state optimization). Together they avoid the false choice between fast-but-wasteful (request-only scaling) and efficient-but-slow (resource-only scaling).

Key design properties

1. Coupled, not independent — vertical output is horizontal input.
2. Different cadences — horizontal is fast (seconds), vertical is slow (tens of seconds) to prevent interference.
3. Asymmetric in both axes — scale-up is aggressive, scale-down is conservative (see patterns/asymmetric-aggressive-up-conservative-down-autoscaling).

Known uses

System	Horizontal signal	Vertical signal	Context
APA	Active concurrent requests (5s)	CPU/GPU util + latency + GPU FLOPS (30s)	Custom Model Serving
systems/databricks-serverless-autoscaler	Query load	OOM detection → larger VM	Serverless Spark

Seen in

• sources/2026-06-10-databricks-ai-serving-platform-that-adapts-to-your-model