Prof. Helena Larsson
Workshop Chair
KTH Royal Institute of Technology
Co-Located Workshop · CN-CSA · 2025
October 20, 2025 · Moscone Center West · San Francisco, USA
CN-CSA is a full-day workshop dedicated to the systems engineering behind consumer-scale marketplaces, search, and recommendation surfaces. It pairs academic researchers in distributed systems with senior industry engineers running production platforms at hundreds of millions of monthly users. The 2025 edition focused on cloud-native patterns that combine retrieval, ranking, and LLM-based reasoning under strict tail-latency and cost budgets, and produced a community-authored short report distributed alongside the main proceedings.
Prof. Helena Larsson
Workshop Chair
KTH Royal Institute of Technology
Dr. Mei Hwang
Workshop Co-Chair
National University of Singapore
Srikanth Jonnakuti
Industry Co-Chair
Realtor.com (News Corp)
Submissions were managed on HotCRP under a double-blind protocol. Each paper received three independent reviews from the 12-member programme committee, followed by online discussion. Final accept/reject decisions were made jointly by the three workshop co-chairs (Larsson, Hwang, Jonnakuti) on the basis of the reviews and discussion.
VP Infrastructure, Google · UC Berkeley
"Beyond CAP: Operating Stateful Consumer Services at Planet Scale"
Latency-Aware Hybrid Retrieval for Marketplace Search
J. Park, A. Singh, M. Ribeiro
Stanford University · Fastly
Prefix-Cache Sharing Across Tenants in Multi-Model Inference Fleets
H. Brooks, L. Zhao, K. Iyer
Anthropic · OpenAI · Google Cloud
Cost-Aware Autoscaling for Recommendation Serving at Consumer Marketplaces
D. Alvarez, J. van der Meer
Booking.com
Evaluation Harnesses for Conversational Search in Real-Estate
A. Sharma, P. Ramanathan, J. Park
Zillow Group · ACM SIGAI · Stanford University
We report on a production evaluation harness used to assess conversational search quality on a U.S. real-estate marketplace serving tens of millions of monthly users. The harness combines offline LLM-judge scoring against a curated 12k-query gold set with an online interleaving framework that compares candidate retrieval and ranking pipelines on live traffic under a strict 200ms p95 budget. We describe the corpus construction, the bias-mitigation procedure for the LLM judge, the variance-reduction techniques used to keep online experiments tractable at our traffic volume, and the operational guardrails (per-tenant rate limits, cost ceilings, and rollback hooks) that allowed the harness to be run continuously in production. The framework has been used to gate every conversational-search release on the platform since Q1 2025, and we summarise the categories of regression it has caught that offline metrics alone missed. An anonymised replication package (gold-set schema, judge prompts, interleaving simulator, and evaluation notebooks) is published alongside the paper.
Guardrails for Listing-Generated Content in Two-Sided Marketplaces
M. Chen, F. Okafor
Airbnb
Multi-Region Personalization with Bounded Staleness
A. Krishnan, R. Iyer
Netflix · Microsoft Azure AI
Trust-and-Safety Pipelines for AI-Assisted Listings
E. Marquez, T. Novak
Red Hat · Amazon Web Services
Edge-Cached Embeddings for Sub-100ms Retrieval
O. Reid, M. Ribeiro
Cloudflare · Fastly
Prof. Andre Dupont
EPFL
Dr. Maya Patel
Amazon Web Services
Sara Okonkwo
Google Cloud
Dr. Rajeev Iyer
Microsoft Azure AI
Lin Zhao
OpenAI
Hannah Brooks
Anthropic
Aditi Krishnan
Netflix
Olivia Reid
Cloudflare
Karthik Iyer
Google Cloud
Prof. Linnea Bergstrom
Chalmers University of Technology
Dr. Rohan Mehta
IIT Bombay
Prof. Maria Chen
University of Toronto
@proceedings{cn_csa_2025,
title = {Proceedings of the Workshop on Cloud-Native Systems for Consumer-Scale Applications (CN-CSA 2025)},
booktitle = {Co-located with Proceedings of the 7th International Conference on Cloud, IoT & Agentic AI (CIOTP 2025)},
year = {2025},
address = {San Francisco, USA},
publisher = {IEEE}
}