---
last_verified: 2026-03-14
description: ADB-S auto-scaling behavior, timing, and sizing implications.
---

auto_scaling:
  notes: "OCPU is legacy billing. ECPU is now default. 1 OCPU ≈ 4 ECPUs."
  ocpu_scaling:
    trigger: "CPU demand exceeds current allocation"
    max_scale: "3x base OCPUs/ECPUs (up to 128 OCPU / 512 ECPU maximum)"
    ramp_up_time: "2-3 minutes to begin scaling"
    ramp_down_time: "After sustained low utilization (5-10 minutes)"
    billing: "Only billed during active scaling period, per OCPU per second"

  storage_scaling:
    trigger: "Storage usage approaches allocated storage"
    behavior: "Auto-expands storage allocation"
    billing: "Billed for allocated storage"

sizing_recommendations:
  base_ocpu_strategy: |
    Size base OCPUs for P75 utilization, not average.
    Auto-scaling handles P95+ bursts, but the 2-3 minute ramp-up
    means very short spikes may complete before scaling activates.

  guidelines:
    - scenario: "Steady OLTP workload"
      base: "P75 utilization"
      auto_scale: "Enable, set max to P95 or 2x base"
    - scenario: "Highly variable workload (2x+ peak/baseline)"
      base: "P50 utilization"
      auto_scale: "Enable, set max to 3x base"
    - scenario: "Batch processing with known peak times"
      base: "Baseline requirement"
      auto_scale: "Enable, supplement with scheduled scaling if available"
    - scenario: "Dev/test"
      base: "1-2 OCPUs"
      auto_scale: "Disable to control costs"

connection_impact:
  connections_per_ocpu: 300
  example: |
    Base 8 OCPUs → 2,400 connections
    Auto-scale to 16 OCPUs → 4,800 connections
    Connection limits scale with OCPUs automatically

performance_during_scaling:
  notes:
    - "Active sessions are not interrupted during scale-up"
    - "New connections get additional capacity immediately"
    - "Scale-down uses connection draining — active sessions complete normally"
    - "Brief performance dip possible during scaling transition"

gotchas:
  - "Short bursts (<3 min) may complete before auto-scaling activates"
  - "Very rapid burst from 0 to max can cause connection queuing"
  - "Auto-scaling cost can surprise if workload is consistently above base"
  - "Monitor auto-scaling events to right-size base OCPUs over time"