Application Refactoring
Moonshot AI needed a more stable and elastic way to preprocess TB-scale multimodal training data for Kimi. Its self-managed Ray and Spark environment struggled with large-cluster stability, could not scale bursty short jobs efficiently, and lacked the job-level observability and priority controls needed by multiple internal teams sharing the same preprocessing estate.
Reported impact
Moonshot AI raised burst-task resource utilization 3x, cut overall compute cost by 45%, reached 99.95% stability for TB-scale single-job preprocessing, and improved observability so high-priority model-data jobs can be scheduled and troubleshot faster.
Moonshot AI rebuilt Kimi's data-preprocessing pipeline on Alibaba Cloud around Alibaba Cloud Container Service for Kubernetes (ACK) and managed Ray and Spark tooling. Text, image, audio, and video corpora are processed by Managed KubeRay and ack-spark-operator workloads inside ACK instead of self-maintained clusters. ACK schedules jobs across ECS capacity and Container Compute Service (ACS), including ACS BestEffort pools for short, fault-tolerant bursts, while Object Storage Service (OSS) provides the shared storage and recovery layer for datasets and event logs. Managed Service for Prometheus, Spark/Ray History Server, and Log Service (SLS) add job-level metrics, dashboards, and centralized logs so teams can watch progress, resource usage, and failures in real time. The result is a more elastic preprocessing platform that improves stability, lowers cost, and gives Moonshot tighter control over Kimi model-data workloads.
Multimodal training data is written to OSS, then Spark and Ray jobs launched through ack-spark-operator and Managed KubeRay run on ACK across ECS and ACS capacity. Metrics flow into Prometheus, execution history and event logs are written back to OSS and History Server, and ECS and ACS pod logs are centralized in SLS so teams can monitor progress and tune job priority or capacity.
ACK's fully managed operations solved large-cluster stability issues and delivered extreme on-demand elasticity with diverse compute options. (original: 通过 ACK 的全托管免运维，解决了大规模集群的稳定性问题。同时充分发挥 ACK 在极致弹性和多样算力方面的关键能力，实现了任务的按需极致弹性伸缩。)
Moonshot AI needed cost-effective and elastic CPU and GPU capacity for model-data preprocessing so it could speed up Kimi's training iteration. (original: 月之暗面作为基础模型创业公司，为了加速大模型的训练迭代，在模型数据预处理方面，需要高性价比、弹性灵活的 CPU 和 GPU 算力满足模型迭代的业务需求。)
Moonshot AI used ACS BestEffort for many jobs, which both kept execution efficient and reduced compute cost. (original: 在部分任务场景中，月之暗面大量采用了 ACS BestEffort 算力，不仅确保了任务的高效执行，还有效地节省了计算成本。)
Under the original architecture, preprocessing ran on self-built Ray and Spark frameworks, but TB-scale jobs suffered from poor large-cluster stability. (original: 在用户原有的架构下，数据预处理任务均基于自建的 Ray 和 Spark 框架开展... 大规模 Spark/Ray 集群稳定性较差，难以保障 TB 级数据预处理任务的稳定执行。)
With the Alibaba Cloud architecture, single-task TB-scale processing stability reached 99.95%, burst-task resource utilization improved 3x, and overall compute cost fell 45%. (original: 单任务 TB 级数据处理稳定性达到 99.95%... 突发性任务资源利用率提升 3 倍，综合计算成本降低 45%。)
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
“联系地址 北京市海淀区知春路76号京东科技大厦1栋13层。”
Related
Alibaba Cloud Sample Jul 28, 2026 Zhuiyin Technology (追因科技) Zhuiyin Technology had already built an agent for cross-border ecommerce analysis, but once the agent was exposed to customers, the team discovered three hard problems: results were hard to judge objectively, reasoning… inference quality improved 20% AgentLoop · Trekin Agent Alibaba Cloud Sample Archived · Jul 21, 2026 Go8 Technology Inc. As Go8 Technology expanded its portfolio of digital solutions, it required a scalable, reliable, and cost-efficient cloud platform to support growing customer demand and increasingly sophisticated workloads. The company… Reduced application latency from 100–250ms to 10–50ms through deployment on local Philippine infrastructure Application Load Balancer (ALB) · Container Service for Kubernetes (ACK) · Content Delivery Network (CDN) Alibaba Cloud Sample Jun 9, 2026 AsiaInfo Technologies (亚信科技) AsiaInfo had to raise software-engineering productivity while maintaining large, long-lived telecom BSS and OSS systems whose logic spanned multiple generations from 2G to 5G. The codebase and architecture were highly… average end-to-end efficiency from requirements to testing improved 8.6% Qoder · Qoder CN 