CONCEPT Cited by 1 source
Hot-node data amplification¶
Hot-node data amplification is the failure mode where consistent hashing + power-law traffic distribution causes a small number of nodes to receive orders of magnitude more data than their peers, compounded by data multiplication during redistribution — multiple upstream instances all routing aggregators for the same hot key to a single owner.
The mechanism¶
- Power-law traffic: popular services (authentication, recommendations) are called by hundreds of upstream services, generating 100× more flow records than typical services.
- Consistent hashing routes by destination: all flow records for a given destination route to the same "owner" instance.
- Data amplification during redistribution: if the hot service has flow records spread across 10 instances (because many clients call it), all 10 instances route their aggregators for that destination to one owner — the volume multiplies at the collection point.
- GC vicious cycle: the hot instance allocates objects faster than it can process → memory pressure → frequent GC pauses → GC consumes CPU → processing slows → memory pressure increases → instance goes DOWN → cascading as load redistributes.
Netflix observed instances handling 100× the flow records of their peers, with GC consuming more CPU than business logic. (Source: sources/2026-07-13-netflix-building-service-topology-at-scale-architecture-challenges)
Why it's worse than simple hot keys¶
Simple hot-key problems (one key gets more traffic) affect one shard. Data amplification is worse because:
- The hot instance receives aggregators from many upstream instances, not just raw data from one source.
- The amplification ratio scales with cluster size (more upstream instances = more aggregators converging on the owner).
- The failure cascades — when the hot instance dies, its load redistributes and creates new hot instances.
Solutions¶
- patterns/graduated-redistribution-for-skewed-data — multi-stage pipelines where each stage re-hashes on different keys, spreading load across multiple distribution points.
- Pre-aggregation at upstream stages (compress before routing).
- Splitting enrichment (I/O-heavy) from resolution (CPU-heavy) into separate stages.
Seen in¶
- sources/2026-07-13-netflix-building-service-topology-at-scale-architecture-challenges — the most severe production issue Netflix faced building Service Topology. Solved by the three-stage pipeline with graduated redistribution + SSE replacing gRPC.
Related¶
- concepts/hotspot — the general concept
- concepts/consistent-hashing — the distribution mechanism that creates the concentration
- concepts/garbage-collection — the JVM resource that amplification exhausts
- patterns/three-stage-flow-aggregation-pipeline — the pattern that solves it
- systems/netflix-service-topology — the production instance