Aurora Serverless v2 will not scale below ~2.5 ACU at 5.6% CPU and 0.5 average active sessions — what determines the floor?

Question

**Setup**
Aurora MySQL 8.0.mysql_aurora.3.10.3 Serverless v2, single writer, no replicas, 0.5–64 ACU, Backtrack on with a 24h window, PI at 465-day retention, engine-default parameter group with no user modifications.                                                
                                                                                                                                     
**Behaviour:**
Averages 6.96 ACU at 5.66% CPU. Lowest ever observed in 14 days is 2.50; typical overnight floor 4.4–5.0 during zero traffic. ACU correlates +0.004 with connections and +0.366 with CPU. At 00:00 UTC with CPU at 2.14%, capacity was still 4.43 and didn't fall across several consecutive quiet hours.                                                                                
                                                                                                                                     
**Metrics table:**
Rhe full set: CPU, AAS, connections, cache hit ratio, query rate, network, volume size, ACU percentiles, top wait event.                                                                                                                             
                                                                                                                                     
**Ruled out:**
connections; pinned parameters (explicitly noting "I know a pinned max_connections can floor Serverless v2 capacity —  that is not the case here", which pre-empts the most common first answer); buffer pool sized to dataset; inefficient queries.      
                                                                                                                                     
**Quesitions:**                                  
                                                                                                                                     
  1. What determines the scale-down floor when CPU, connections and buffer pool are all non-binding? Any observable metric?          
  2. Does Backtrack impose a minimum capacity or hold proportional memory?                                                           
  3. Is there documented behaviour where sustained time at max_capacity elevates the floor afterwards, and what's the decay?         
  4. Does PI with long retention contribute?                                                                                         
  5. Does max_capacity itself affect scale-down granularity?

Riku_Kobayashi · Answer

Hello.

Aurora Serverless v2 scales by evaluating aggregate resource usage—collectively termed "load"—that encompasses not only CPU but also memory, network, internal database processing, and Aurora management tasks.  
AWS explicitly states that capacity may not always scale down to the minimum level, even when there is no load.  
https://docs.aws.amazon.com/AmazonRDS/latest/AuroraUserGuide/aurora-serverless-v2.how-it-works.html

If the features described in the following document are enabled, resource usage may increase, and the system may fail to scale down.  
https://docs.aws.amazon.com/AmazonRDS/latest/AuroraUserGuide/aurora-serverless-v2.setting-capacity.html#aurora-serverless-v2.troubleshooting  
> Certain features can increase resource usage and prevent the database from scaling down to minimum capacity. These features include the following:  
> * Aurora global databases  
> * Exporting CloudWatch Logs  
> * Enabling pg_audit on Aurora PostgreSQL–compatible DB clusters  
> * Enhanced Monitoring  
> * Performance Insights

Aurora Serverless v2 will not scale below ~2.5 ACU at 5.6% CPU and 0.5 average active sessions — what determines the floor?

Add your answer

Relevant content