Since the last decade, public cloud platforms are rapidly becoming de-facto computing platform for our society. To support the wide range of users and their diverse applications, public cloud platforms started to offer the same VMs under many purchasing options that differ across their cost, performance, availability, and time commitments. Popular purchasing options include on-demand, reserved, and transient VM types. Reserved VMs require long time commitments, whereas users can acquire and release the on-demand (and transient) VMs at any time. While transient VMs cost significantly less than on-demand VMs, platforms may revoke them at any time. In general, the stronger the commitment, i.e., longer and less flexible, the lower the price. However, longer and less flexible time commitments can increase cloud costs for users if future workloads cannot utilize the VMs they committed to buying. Interestingly, this wide range of purchasing options provide opportunities for cost savings. However, large cloud customers often find it challenging to choose the right mix of purchasing options to minimize their long-term costs while retaining the ability to adjust their capacity up and down in response to workload variations. Thus, optimizing the cloud costs requires users to select a mix of VM purchasing options based on their short- and long-term expectation of workload utilization. Notably, hybrid clouds combine multiple VM purchasing options or private clusters with public cloud VMs to optimize the cloud costs based on their workload expectations. In this thesis, we address the challenge of choosing a mix of different VM purchasing options in the context of large cloud customers and thereby optimizing their cloud costs. To this end, we make the following contributions: (i) design and implement a container orchestration platform (using Kubernetes) to optimize the cost of executing mixed interactive and batch workloads on cloud platforms using on-demand and transient VMs, (ii) develop simple analytical models for different straggler mitigation techniques to better understand the cost of synchronization in distributed machine learning workloads and compare their cost and performance on on-demand and transient VMs, (iii) design multiple policies to optimize long-term cloud costs by selecting a mix of VM purchasing options based on short- and long-term expectations of workload utilization (with no job waiting), (iv) introduce the concept of waiting policy for cloud-enabled schedulers, and show that provisioning long-term resources (e.g., reserved VMs) to optimize the cloud costs is dependent on it, and (v) design and implement speculative execution and ML-based waiting time predictions (for waiting policies) to show that optimizing job waiting in the cloud is possible without accurate job runtime predictions.