Stop Pretending to Be Busy: A Case for Serverless Paradigms in Co-Located Batch Workloads (Operational Systems)

High resource utilization is significant for cloud vendors. To achieve this, a common practice is to co-locate low-priority batch workloads (typically Spark-based analytics) with high-priority online services, while strictly maintaining Service Level Objectives (SLOs). This paper presents an empirical study of co-location and overcommitment in a production-scale datacenter. Specifically, within Ant Group, online services utilize only 22.0% of available CPU resources. By carefully overcommitting resources to deploy batch workloads, the system harvests an additional 26.8% of CPU capacity.

Despite this increased density, we observe that batch workloads remain inefficient, with a useful computation ratio of only 67%. We identify the root causes of this low "effective utilization" as four types of idleness: (1) slot idle, arising from coarse-grained resource management in Spark; (2) gap idle, caused by hardware heterogeneity and interference; and (3/4) start/stop idles, resulting from the high latency of launching and destroying analytic instances. To address these inefficiencies, we propose Quark, a novel framework that integrates serverless paradigms into batch analytics. Quark eliminates these idles through fine-grained resource allocation, heterogeneity and skew-aware scheduling, and rapid instance provisioning. Experimental results show that Quark increases cluster utilization by about 37.37% and reduces the proportion of long-tail jobs from 15% to 2%. Quark has been deployed at scale within Ant Group, processing 350,000 offline query jobs daily across a deployment footprint of 600,000 CPU cores, processing between 7,500 TB and 10,000 TB of data daily, and saving more than 100,000 CPU cores.