Choosing Node-Based or Serverless Before Traffic Patterns Decide for You

Decision drivers

Traffic predictability and variability over a typical week
Average payload size per request (drives ECPU cost directly)
Whether Reserved Node discounting or Global Datastore cross-region replication is a real requirement
Tolerance for the Serverless minimum-metering floor on a low-traffic cache

Production premise

Node-based ElastiCache bills per node-hour 24/7 regardless of traffic, with access to Reserved Node discounts and Global Datastore. Serverless bills on two independent meters — GB-hours of data stored and ECPUs (roughly 1 ECPU per KB transferred, more for complex commands like sorted-set operations) — with no node to size, but also no Reserved Node pricing and no Global Datastore.

Failure mode & inflection point

Serverless has a real floor: even an empty or near-empty cache is metered continuously (100 MB minimum for Valkey, versus 1 GB for Redis OSS/Memcached — a meaningfully lower floor on Valkey specifically). At sustained high request volume, ECPU costs can exceed what an equivalently-sized node-based cluster would cost, because ECPU billing scales with data transferred per request, not with a flat capacity reservation — a workload with large payloads per GET/SET (say, serialized objects in the tens of KB) burns ECPUs fast even at modest request rates. Conversely, node-based pricing is a fixed cost whether the cache is busy or idle 90% of the day.

Production guardrail

Default to node-based for anything that's past initial ramp-up and running with roughly steady, predictable traffic — which describes most production caching layers once they've settled — since a fixed 24/7 node cost beats ECPU metering once request volume is consistent, and node-based is the only path to Reserved Node discounts or Global Datastore. Reach for Serverless instead when traffic is genuinely bursty or unpredictable (pre-launch, spiky batch/event-driven workloads, a fleet of low-traffic dev/staging caches where the 100 MB Valkey floor keeps cost trivial) — in those cases the removed capacity-planning overhead outweighs the per-request premium. If Reserved Nodes or Global Datastore are a real requirement, that alone settles it toward node-based regardless of traffic shape.