Retrieval-Augmented Generation (RAG) and other AI/ML workloads rely on vector databases (DBs) for efficient analysis of unstructured data. However, cluster (or serverful ) vector DB architectures, such as Milvus, lack the elasticity to handle high workload fluctuations, sparsity, and burstiness. Serverless vector DBs-- i.e., vector DBs built on top of cloud functions--have emerged as a promising alternative architecture, but they are still in their infancy. This paper presents the first experimental study comparing data partitioning strategies in vector DBs built atop stateless Function-as-a-Service (FaaS). Through extensive benchmarks, we reveal key limitations of clustering-based data partitioning when applied to dynamic datasets ( e.g. , complexity, load balancing). We then evaluate a block-based alternative that addresses such limitations ( e.g. , up to 5.8× faster data partitioning, up to 63% lower costs, similar querying times). Moreover, our results show that a stateless serverless vector DB using block-based data partitioning achieves competitive performance with Milvus in several aspects ( e.g. , up to 65.6× faster data partitioning, similar recall), while reducing costs for sparse workloads (up to 99%). Our empirical insights aim to guide the design of next-generation serverless vector DBs.

The elasticity of the Cloud is very appealing for processing large scientific data. However, enormous volumes of unstructured research data, totaling petabytes, remain untapped in data repositories due to the lack of efficient parallel data access. Even-sized partitioning of these data to enable its parallel processing requires a complete re-write to storage, becoming prohibitively expensive for high volumes. In this article we present Dataplug, an extensible framework that enables fine-grained parallel data access to unstructured scientific data in object storage. Dataplug employs read-only, format-aware indexing, allowing to define dynamically-sized partitions using various partitioning strategies. This approach avoids writing the partitioned dataset back to storage, enabling distributed workers to fetch data partitions on-the-fly directly from large data blobs, efficiently leveraging the high bandwidth capability of object storage. Validations on genomic (FASTQGZip) and geospatial (LiDAR) data formats demonstrate that Dataplug considerably lowers pre-processing compute costs (between 65.5% — 71.31% less) without imposing significant overheads.

Serverless data analytics generate a large amount of intermediate data during computation stages. However, serverless functions, which are short-lived and lack direct communication, face significant challenges in managing this data effectively. The traditional approach of using object storage to carry the data proves to be slow and costly, as it involves constant movement of data back and forth. Although specialized ephemeral storage solutions have been developed to address this issue, they fail to tackle the fundamental challenge of minimizing data movements. This work focuses on incorporating near-data computation into an ephemeral storage system to reduce the volume of transferred data in serverless analytics. We present Glider with the aim to enhance communication between serverless compute stages, allowing data to smoothly "glide" through the processing pipeline instead of bouncing between different services. Glider achieves this by leveraging stateful near-data execution of complex data-bound operations and an efficient I/O streaming interface. Under evaluation, it reduces data transfers by up to 99.7%, improves storage utilization by up to 99.8%, and enhances performance by up to 2.7×. In sum, Glider improves serverless data analytics by optimizing data movement, streamlining processing, and avoiding redundant transfers.

Serverless computing greatly simplifies the use of cloud resources. In particular, Function-as-a-Service (FaaS) platforms enable programmers to develop applications as individual functions that can run and scale independently. Unfortunately, applications that require fine-grained support for mutable state and synchronization, such as machine learning (ML) and scientific computing, are notoriously hard to build with this new paradigm. In this work, we aim at bridging this gap. We present Crucial , a system to program highly-parallel stateful serverless applications. Crucial retains the simplicity of serverless computing. It is built upon the key insight that FaaS resembles to concurrent programming at the scale of a datacenter. Accordingly, a distributed shared memory layer is the natural answer to the needs for fine-grained state management and synchronization. Crucial allows to port effortlessly a multi-threaded code base to serverless, where it can benefit from the scalability and pay-per-use model of FaaS platforms. We validate Crucial with the help of micro-benchmarks and by considering various stateful applications. Beyond classical parallel tasks (e.g., a Monte Carlo simulation), these applications include representative ML algorithms such as k -means and logistic regression. Our evaluation shows that Crucial obtains superior or comparable performance to Apache Spark at similar cost (18%–40% faster). We also use Crucial to port (part of) a state-of-the-art multi-threaded ML library to serverless. The ported application is up to 30% faster than with a dedicated high-end server. Finally, we attest that Crucial can rival in performance with a single-machine, multi-threaded implementation of a complex coordination problem. Overall, Crucial delivers all these benefits with less than 6% of changes in the code bases of the evaluated applications.

Serverless computing has seen a myriad of work exploring its potential. Some systems tackle Function-as-a-Service (FaaS) properties on automatic elasticity and scale to run highly-parallel computing jobs. However, they focus on specific platforms and convey that their ideas can be extrapolated to any FaaS runtime. An important question arises: do all FaaS platforms fit parallel computations? In this paper, we argue that not all of them provide the necessary means to host highly-parallel applications. To validate our hypothesis, we create a comparative framework and categorize the architectures of four cloud FaaS offerings, emphasizing parallel performance. We attest and extend this description with an empirical experiment that consists in plotting in deep detail the evolution of a parallel computing job on each service. The analysis of our results evinces that FaaS is not inherently good for parallel computations and architectural differences across platforms are decisive to categorize their performance. A key insight is the importance of virtualization technologies and the scheduling approach of FaaS platforms. Parallelism improves with lighter virtualization and proactive scheduling due to finer resource allocation and faster elasticity. This causes some platforms like AWS and IBM to perform well for highly-parallel computations, while others such as Azure present difficulties to achieve the required parallelism degree. Consequently, the information in this paper becomes of special interest to help users choose the most adequate infrastructure for their parallel applications.

Serverless computing is an emerging paradigm that greatly simplifies the usage of cloud resources and suits well to many tasks. Most notably, Function-as-a-Service (FaaS) enables programmers to develop cloud applications as individual functions that can run and scale independently. Yet, due to the disaggregation of storage and compute resources in FaaS, applications that require fine-grained support for mutable state and synchronization, such as machine learning and scientific computing, are hard to build.