During the last decade, the proliferation of smartphones, social media and streaming services has provoked an explosion of multimedia data. This large amount of image and video sources combined with more powerful and inexpensive computational capabilities brought by the cloud computing paradigm has facilitated the rapid growth of new machine learning models capable of extracting information faster and more accurately. However, the complexity to develop machine learning models has also grown, involving multiple steps, from the acquisition and preparation of data to the training, evaluation and deployment of models. To alleviate this, the leading database providers have started to integrate the predictive capabilities of machine learning directly into their systems. This new approach is known as in-database machine learning, and it brings new interesting properties such as the exploitation of the inherent relational structure of data or the preservation of its privacy and integrity since the inference occurs directly where the data lives. In this work, we present a cloud-native approach to perform in-database machine learning. We have extended SQLFlow, a bridge between SQL engines and machine learning toolkits to support models trained to solve image recognition tasks over image datasets, which meta-information is persisted on a relational database. Furthermore, we have encapsulated the definition of machine learning models on cloud-native workflows that are able to exploit the GPU resources available in a Kubernetes environment. Our research evaluates the scalability of the proposed system regarding the total execution time and GPU utilization. Besides, we are interested in exploring the design of optimized machine learning query plans, where the goal is to choose the optimal among multiple models that cover a range of specific classes to predict from attending its accuracy and execution cost. For that purpose, we have implemented a model repository containing different model variations and evaluated different strategies to optimize the model selection. Our experiments show that optimizing the model selection will lead to more accurate and faster results, especially when a query covers a high number of classes and the number of models that are able to answer them is limited.

At the moment we are witnessing the maturation of distributed streaming dataflow systems whose use-cases have departed from the mere analysis of streaming windows and complex-event processing, as they now extend to cloud applications, workflows and even e-commerce. The state of streaming operators has been so far hidden from external applications. In this thesis it is argued that exposing this internal state to outside applications by making it queryable, opens the road for novel use-cases. To this end, we introduce S-Query: a system and reference architecture where the state of stream operators can be queried - either live or through snapshots, achieving different isolation levels. It is shown how this can be implemented in an existing open-source stream processor and how queryable state can affect the performance of such a system. Our experimental evaluation suggests that snapshot configuration adds only up to 5ms in latency in the 99.99th percentile and no increase in 0-90th percentiles.

Major advances in the fault tolerance of distributed stream processing systems provided the systems with the capacity to produce strictly consistent results under failures. Consistent fault tolerance has been one of the catalysts fueling the maturity of streaming systems and boosting their widespread adoption not only for analytics use cases, but also as a platform for running applications, such as microservices and stateful functions. The most common fault tolerance strategy that many modern streaming systems adopt is an adaptation of Chandy-Lamport's distributed snapshots protocol that is based on global coordinated checkpoints. Given the importance of fault tolerance and the impact of coordinated checkpoints in the stream processing space, it is surprising that no other fault tolerance algorithm has been considered as an alternative. In this paper we bring together and benchmark three seminal fault tolerance algorithms from the distributed systems literature: the coordinated, uncoordinated, and communication-induced checkpoint algorithms. We evaluate their behavior in terms of runtime performance and failure recovery on a variety of streaming workloads including pipeline, scatter-gather, and cyclical topologies. We benchmark the algorithms on a novel streaming system, FERDiS, built from scratch as an extensible benchmarking framework. The experiment results show that the coordinated approach that state of the art streaming systems adopt is in many cases optimal, but with some exceptions where it is outperformed in both runtime performance and failure recovery by the uncoordinated and communication-induced algorithms. The cause of the differences in performance, both at runtime and during recovery, was found to be strongly related to the characteristics of the input stream(s) and the query graph. On the tested cyclic topology, the coordinated checkpointing algorithm could not be applied as by its design it will deadlock on such query graphs. The natural cycle support of the uncoordinated and communication -induced proved advantageous here and we observed high susceptibility to the unbounded domino effect of the uncoordinated algorithm which the communication-induced was not affected by, indicating the unsuitability of uncoordinated checkpointing in cyclic stream processing.

Lately, serverless computing has gained much attention. Function-as-a-Service (FaaS) is the most prominent serverless model. The FaaS model allows cloud users to write simple functions, and the cloud provider takes care of the deployment, main- tenance, and scalability of those functions. However, FaaS can be improved upon significantly. Existing FaaS services do not solve some significant and challenging cloud computing problems regarding state and communication, such as function- to-function calls and transactions. Work has been done to improve this early it- eration of serverless computing by incorporating state in the equation; leading to Stateful Function-as-a-Service (SFaaS). However, transactions on SFaaS remain an open problem, even though they are crucial to allow more use cases to operate using SFaaS. This thesis examines existing SFaaS systems and their data correctness guaran- tees. Based on existing SFaaS systems, this thesis proposes a programming model for both serializable transactions and sagas orchestrations using coordinator functions. The programming model is implemented in Apache Flink StateFun. The implemen- tation is based on existing research on traditional database technology. The implementation’s performance is evaluated using an enhanced version of the Yahoo! Cloud Systems Benchmark (YCSB); YCSB is extended to include a trans- actional operation. The implementation reaches 1840 sagas workflows per second and 1200 serializable transactions at sub-200ms latencies on 5 workers with 2 CPUs each.