Back office ops · Production

Nextdoor CoreML team achieves 4x latency reduction and 3x throughput increase by tuning ML inference in shared hosting environments

The problem

Running ML inference in shared hosting environments (ECS, K8s) introduces unobvious pitfalls that significantly impact latency and throughput.

First attempt

Nextdoor's ML team experienced latency and throughput degradation traced to poor request queue management and suboptimal OpenMP parameter configuration before discovering and resolving the root causes.

Workflow diagram · grounded in source
1
ML inference request arrives
trigger
“Running a ML inference layer in a shared hosting environment (ECS, K8s, etc.)”
2
Load balancer routes request
routing
“Why your load balancing algorithm matters”
3
Request queue timeout management
validation
“The importance of request queue timeouts for service recovery”
4
OpenMP parameter tuning
integration
“The main points of concern are request queue management and OpenMP parameter tuning”
5
ML inference result delivered
output
“ML infrastructure for a variety of production ML use cases at Nextdoor, such as feed, notifications and ads”
Reported outcome

After addressing request queue management and OpenMP parameter tuning, Nextdoor achieved a factor of 4 latency reduction, 3x throughput increase, and CPU utilization improvement from 10% to 50% while maintaining model performance.

Reported metrics
Latency reductionfactor of 4
Throughput increase3x
CPU utilization improvementCPU 10% -> 50%
real-time ML microservices in production30+
Show all 5 reported metrics
latency reductionfactor of 4
throughput increase3x
CPU utilization improvementCPU 10% -> 50%
real-time ML microservices in production30+
years of ML operationalisation experience6 years
Reported stack
ECSK8sOpenMP
Source
https://engblog.nextdoor.com/running-ml-inference-services-in-shared-hosting-environments-6176b39bc9b7
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

After addressing request queue management and OpenMP parameter tuning, Nextdoor achieved a factor of 4 latency reduction, 3x throughput increase, and CPU utilization improvement from 10% to 50% while maintaining model…

What tools did this team use?

ECS, K8s, OpenMP.

What results were reported?

Latency reduction: factor of 4; Throughput increase: 3x; CPU utilization improvement: CPU 10% -> 50%; real-time ML microservices in production: 30+ (source-reported, not independently verified).

What failed first in this deployment?

Nextdoor's ML team experienced latency and throughput degradation traced to poor request queue management and suboptimal OpenMP parameter configuration before discovering and resolving the root causes.

How is this back office ops AI workflow structured?

ML inference request arrives → Load balancer routes request → Request queue timeout management → OpenMP parameter tuning → ML inference result delivered.