AWS re:Invent 2025 - How Netflix Shapes our Fleet for Efficiency and Reliability (IND387)

Balancing Efficiency and Reliability at Netflix

Understanding Efficiency and Reliability

• Efficiency is about maximizing the business value of workloads while minimizing the cost of running them.
• Reliability is about ensuring predictable performance, infrequent failures, and fast recovery when failures do occur.
• Efficiency and reliability are complementary - Netflix views them as a single construct, not separate goals.
• Key efficiency metrics include cost, resource usage, and the risk-adjusted net value of workloads.
• Key reliability metrics include mean time between failures, mean time to recovery, and the blast radius of failures.

Managing the Supply of Compute Resources

• Capacity planning is critical to allocate compute resources efficiently across Netflix's fleet.
• Netflix models both stateful (e.g. databases) and stateless (e.g. microservices) workloads differently for capacity planning.
• Reserving the right mix of reserved and on-demand capacity is key to balancing efficiency and cost.
• Monitoring hardware availability and characteristics (e.g. generation, shape) is crucial to plan for volatile capacity.
• Netflix uses the concepts of "success buffer" and "failure buffer" to model headroom and plan capacity.
• Buffers vary by workload type and hardware - Netflix tailors them for critical vs. non-critical services.

Understanding Compute Demand

• Profiling workloads to understand CPU, memory, and network resource needs is foundational.
• Observing production workloads is crucial to validate models and understand real-world usage patterns.
• Accounting for service startup times is important when planning for capacity.
• Netflix's microservices architecture creates complex, non-linear call patterns that impact demand estimation.
• Predictable daily and weekly traffic patterns are complemented by unpredictable spikes during major events.

Balancing Supply and Demand

• Fleet shaping - matching workloads to optimal hardware based on performance, cost, and capacity constraints.
• Pre-scaling - proactively scaling the fleet in anticipation of traffic spikes to maintain efficiency and reliability.
• Dynamic traffic shaping - redistributing existing and new traffic across regions to balance load.
• Reactive auto-scaling - rapidly adding or removing capacity in response to demand changes.
• Prioritized load shedding - intelligently shedding non-critical traffic to protect core functionality.

Key Takeaways

• Efficiency and reliability must be managed holistically, not as separate goals.
• Detailed modeling and observation of both supply and demand is critical.
• Proactive and reactive techniques are needed to balance efficiency and reliability.
• Compound efficiency wins are possible at Netflix's scale.
• End-to-end traffic management and "math as a safety blanket" are key principles.