BERND KUHLENAll notes ↗The right EFS throughput mode is not the one with the largest headline number. It is the one whose capacity model matches the shape of the workload and whose cost and failure modes the team can operate confidently.
Amazon EFS currently offers Elastic, Provisioned and Bursting throughput. They solve different problems.
Bursting: capacity follows stored data
Bursting is a natural fit when throughput requirements grow with the amount of data in Standard storage and busy periods are followed by enough quiet time to replenish credits. It can be economical and simple, but its sustainable baseline may be surprisingly low for small file systems.
Choose it when the storage footprint and throughput demand genuinely scale together. Avoid treating accumulated credits as permanent capacity. A continuously falling BurstCreditBalance proves that the workload is not in equilibrium.
Elastic: capacity follows activity
Elastic Throughput automatically scales with workload activity and does not use burst credits. AWS recommends it for spiky or unpredictable workloads, particularly when average throughput is a small fraction of the peak. This removes the credit-exhaustion failure mode and reduces capacity planning effort.
The trade-off is commercial rather than architectural: Elastic Throughput is usage-based, so high or poorly governed I/O can increase cost. Monitor MeteredIOBytes, establish budgets and investigate unexpected demand rather than assuming automatic scaling removes the need for governance.
Provisioned: capacity follows an explicit requirement
Provisioned Throughput decouples throughput from file-system size and provides a chosen capacity level. It suits workloads with a known, sustained requirement or a higher average-to-peak ratio. The benefit is predictability; the responsibility is to size and revisit the provisioned amount.
Provisioning too little recreates a capacity constraint. Provisioning too much wastes money. This mode works best when measurements are stable enough to support an explicit commitment.
A practical selection framework
- Use Bursting when storage size is a meaningful predictor of throughput, quiet periods replenish credits and credit monitoring is part of normal operations.
- Use Elastic when demand is irregular, peaks are difficult to forecast or avoiding credit exhaustion is worth usage-based pricing.
- Use Provisioned when throughput demand is known, relatively sustained and independent of stored-data size.
Migration is a measured change
Before changing mode, capture baseline metrics, expected peak and average demand, current credit trend and application latency. After the change, compare those same signals and review cost. AWS also documents temporary restrictions after switching to or changing Provisioned Throughput, so the runbook should include the applicable timing constraints.
The decision in one sentence
If stored data predicts demand, Bursting may fit. If activity is unpredictable, Elastic is usually the safer operational model. If the requirement is stable and measurable, Provisioned can provide deliberate capacity. Whichever mode you choose, make its limiting resource visible.