Device-cloud collaborative learning trains a combined model using the cloud and distributed edge devices without exposing raw data. Each co-computing round involves: 1) devices fetching the latest model weights from the cloud for local training; 2) the cloud aggregating delta weights from devices to update the model. However, edge devices are geo-distribution and may perform co-computing in different time zones, which introduces three new challenges to existing systems: 1) task execution is delayed until reaching the device’s available time window, the cloud needs an extended period to collect model weights. 2) devices have varying availability, but current task assignment assumes equal available training time, causing task imbalance. 3) massive model weights may arrive at unpredictable times, the cloud must reserve huge resources for peak loads. To fill the gap, we propose OLearning, a geo-distributed production system. Specifically, OLearning 1) applies a two-layer multi-zone design. The first layer performs inter-zone weight aggregation on various zones from the second layer, while the second layer performs intra-zone weight aggregation on selected devices. 2) adds a reward indicator for each task, allowing devices to decide task scheduling orders locally. 3) deploys a shared-task resource cluster for aggregating millions of delta weights over undetermined arrival times. Comprehensive experiments on the public dataset show the effectiveness and robustness of OLearning.