Testing Neki's Ability to Handle Millions of Queries
Neki demonstrates linear scalability by sustaining over 118 million queries per second across 512 shards. This performance is achieved by distributing a simple point-select workload across a massive cluster of Postgres primaries and routers. The results prove that the system can scale throughput proportionally as more shards are added. This capability allows for handling extreme data volumes and request rates without sacrificing stability.
key points
The benchmark used a read-only workload of single-shard point selects by primary key, avoiding complex joins or cross-shard queries. This simplified workload allows the system to isolate traffic to specific shards, maximizing raw throughput.
Scaling from 5 to 512 shards resulted in a linear increase in total QPS, with per-shard performance remaining consistent. This indicates that the system does not suffer from significant overhead or bottlenecks as the cluster size grows.
The infrastructure required for this peak performance included 512 Postgres primaries on r8g.16xlarge instances and 480 Neki routers on 8xlarge instances. This massive hardware fleet managed 1.22 PiB of data with a p99 client latency of 13.95ms.
community discussion
5 Mixed[consensus]
Commenters are skeptical of the high costs and proprietary nature of the Neki benchmark, with several arguing that closed-source software is a dealbreaker for production use. There is a technical debate regarding historical AWS capabilities, specifically whether reading terabytes of data in seconds was possible in 2015. Most agree that while the scale achieved is impressive, the high cache hit rate (87.3%) suggests the results measure cache performance more than raw query execution.
top insight
A significant technical point is raised regarding the trade-off between single-node and distributed databases, suggesting a distributed system needs to fan out to 50-100 nodes to outperform a cache-optimized single-node database due to modern CPU cache line architectures. This provides a critical lens for evaluating whether 'planet-scale' distributed systems are always the most efficient choice for every workload.