Publications

Alloy: Transparent Proxy-Based Coupling Between Stream Processing and Storage

Published in DEBS '26: Proceedings of the 19th ACM International Conference on Distributed and Event-based Systems, 2026

Stream storage and processing workflows support a wide range of applications, from enterprise data lakes to IoT analytics. Streams are ingested and stored in broker clusters, e.g., using Apache Kafka, while processing engines like Apache Flink execute queries. In such a setup, however, streams are read, sent, and (de)serialized in full between storage and processing, even when only a subset is relevant, leading to inefficiencies in networking and computation. We present Alloy, a novel approach for reducing the cost of coupling stream storage and processing. Alloy filters data on storage nodes and sends only what is useful for a specific query to the processing cluster. A distinctive feature of Alloy is its transparency: the data plane uses WebAssembly plugins deployed atop the Envoy proxy, interposing between Apache Flink and Apache Kafka to perform selection and projection without modifying either system. The Alloy control plane analyzes queries as part of the regular Flink submission and configures proxies accordingly. Evaluation using microbenchmarks and Nexmark shows that Alloy drastically reduces network usage while maintaining or reducing CPU utilization, with minimal latency impact.

Recommended citation: Rosinosky, G., Schmitz, D., & Riviere, E. (2026, June). Alloy: Transparent Proxy-Based Coupling Between Stream Processing and Storage. In Proceedings of the 20th ACM International Conference on Distributed and Event-based Systems (pp. 23-36). https://dl.acm.org/doi/full/10.1145/3701717.3734462

Justin: Hybrid CPU/Memory Elastic Scaling for Distributed Stream Processing

Published in IFIP International Conference on Distributed Applications and Interoperable Systems (DAIS) 2025, 2025

Distributed Stream Processing (DSP) engines analyze continuous data via queries expressed as a graph of operators. Auto-scalers adjust the number of parallel instances of these operators to support a target rate. Current auto-scalers couple CPU and memory scaling, allocating resources as one-size-fits-all packages. This contrasts with operators’ high diversity of requirements. We present Justin, an auto-scaler that enables hybrid CPU and memory scaling of DSP operators. Justin monitors both CPU usage and the performance of operators’ storage operations. Its mechanisms enable fine-grain memory allocation for tasks upon a query reconfiguration. The Justin policy identifies individual operators’ memory pressure and decides between adjusting parallelism and/or memory assignment. We implement Justin in Apache Flink, extending the Flink Kubernetes Operator and the DS2 CPU-only auto-scaler. Using the Nexmark benchmark, our evaluation shows that Justin identifies suitable resource allocation in as many or fewer reconfiguration steps as DS2 and supports a target rate with significantly fewer CPU and memory resources.

Recommended citation: Schmitz, D., Rosinosky, G., & Rivière, E. (2025, June). Justin: Hybrid CPU/Memory elastic scaling for distributed stream processing. In IFIP International Conference on Distributed Applications and Interoperable Systems (pp. 102-118). Cham: Springer Nature Switzerland. https://link.springer.com/chapter/10.1007/978-3-031-95728-4_6

Charon: Fine-Grain Resource Allocation and Optimized Task Placement for Multi-Query Distributed Stream Processing

Published in SAC '26: Proceedings of the 41st ACM/SIGAPP Symposium on Applied Computing, 2025

Distributed Stream Processing (DSP) enables the analysis of large volumes of continuous data. A DSP engine generally executes multiple queries on a shared cluster. However, state-of-the-art DSP engines such as Apache Flink provision resources to queries in isolation and as one-size-fits-all units of CPU and memory, despite the variety of their requirements. We propose Charon, a novel approach that dynamically assigns resources to processing elements at fine-grained levels and considers the complete set of colocated queries in its placement strategies, thereby improving consolidation and resource efficiency. Charon uses optimized placement strategies based on a bin-packing optimization algorithm, while respecting the throughput requirements of each query. We implement Charon in Flink and the Flink Kubernetes Operator. We evaluate our approach using queries from the Nexmark multi-query benchmark and demonstrate its effectiveness in improving resource utilization compared to Flink.

Recommended citation: Schmitz, D., Rosinosky, G., & Rivière, E. (2026, March). Charon: Fine-Grain Resource Allocation and Optimized Task Placement for Multi-Query Distributed Stream Processing. In Proceedings of the 41st ACM/SIGAPP Symposium on Applied Computing (pp. 265-273). https://dl.acm.org/doi/abs/10.1145/3748522.3779716

A Tale of Many Streams: Characterizing a Hybrid Batch-Stream Production Workload in Digazu, a Data Lake supported by Apache Kafka and Flink

Published in DEBS '25: Proceedings of the 19th ACM International Conference on Distributed and Event-based Systems, 2025

Many industries rely on analyzing large volumes of combined historical and live data. A data lake facilitates these operations by supporting an integrated data ingestion, storage, replay, and analysis workflow. A modern data lake is distributed and combines a processing engine, able to seamlessly process large volumes of existing data as well as continuous flows of new data, such as Apache Flink, with a storage infrastructure able to ingest and replay this data, such as Apache Kafka. This use of Flink in this setting departs from the commonly agreed model of stream processing queries operating over windows of events, maintaining a bounded and relatively small state per operator. Instead, hybrid batch-stream queries typically process an existing data set in its entirety before updating results with incoming stream data, leading to a large accumulated state. Given the industry’s importance of such usages, understanding their characteristics and how they differ from common assumptions in designing and evaluating stream processing systems is of utmost importance. We present in this paper the analysis of a large-scale hybrid batch-stream workload collected from a production deployment of Digazu, a modern data lake building upon Kafka and Flink. We characterize 142 different sources of data and 129 hybrid batch-stream queries. Our analysis offers valuable insights into the nature of data and queries in typical data lake deployment, which will assist designers of such systems and associated benchmarks.

Recommended citation: Schmitz, D., Berrewaerts, L., Rosinosky, G., Skhiri, S., & Rivière, E. (2025, June). A Tale of Many Streams: Characterizing a Hybrid Batch-Stream Production Workload in Digazu, a Data Lake supported by Apache Kafka and Flink. In Proceedings of the 19th ACM International Conference on Distributed and Event-based Systems (pp. 188-198). https://dl.acm.org/doi/full/10.1145/3701717.3734462

StreamBed: capacity planning for stream processing

Published in DEBS '24: Proceedings of the 18th ACM International Conference on Distributed and Event-based Systems, 2024

StreamBed is a capacity planning system for stream processing. It predicts, ahead of any production deployment, the resources that a query will require to process an incoming data rate sustainably, and the appropriate configuration of these resources. StreamBed builds a capacity planning model by piloting a series of runs of the target query in a small-scale, controlled testbed. We implement StreamBed for the popular Flink DSP engine. Our evaluation with large-scale queries of the Nexmark benchmark demonstrates that StreamBed can effectively and accurately predict capacity requirements for jobs spanning more than 1,000 cores using a testbed of only 48 cores.

Recommended citation: Rosinosky, G., Schmitz, D., & Rivière, E. (2024, June). StreamBed: capacity planning for stream processing. In Proceedings of the 18th ACM International Conference on Distributed and Event-based Systems (pp. 90-102). https://dl.acm.org/doi/abs/10.1145/3629104.3666034

C2B2: a Cloud-native Chaos Benchmarking suite for the Hyperledger Fabric Blockchain

Published in IEEE EDCC, 2022

Stream processing frameworks allow processing massive amounts of data shortly after it is produced, and enable a fast reaction to events in scenarios such as data center monitoring, smart transportation, or telecommunication networks. Many scenarios depend on the fast and reliable processing of incoming data, requiring low end-to-end latencies from the ingest of a new event to the corresponding output. The occurrence of faults jeopardizes these guarantees: Currently- leading high-availability solutions for stream processing such as Spark Streaming or Apache Flink’s implement passive replication through snapshotting, requiring a stop-the-world operation to recover from a failure. Active replication, while incurring higher deployment costs, can overcome these limitations and allow to mask the impact of faults and match stringent end-to-end latency requirements. We present the design, implementation, and evaluation of active replication in the popular Apache Flink platform. Our study explores two alternative designs, a leader-based approach leveraging external services (Kafka and ZooKeeper) and a leaderless implementation leveraging a novel deterministic merging algorithm. Our evaluation using a series of microbenchmarks and a SaaS cloud monitoring scenario on a 37-server cluster show that the actively-replicated Flink can fully mask the impact of faults on end-to-end latency.

Recommended citation: Kassab, A., Riviere, E., Rosinosky, G., Sadre, R., & Tran, V. H. (2022, September). C2B2: a Cloud-native Chaos Benchmarking suite for the Hyperledger Fabric Blockchain. In 2022 18th European Dependable Computing Conference (EDCC) (pp. 89-96). IEEE. http://hdl.handle.net/2078.1/267055

Machine‐as‐a‐Service: Blockchain‐based management and maintenance of industrial appliances

Published in Engineering reports, 2022

Machine-as-a-Service (MaaS) is an emerging service model for industrial appliances. With MaaS, machines are rented instead of being acquired, and their lifecycle is handled by an ecosystem of specialized actors, such as different independent maintenance companies certified for interventions on specific hardware. As the number of actors, clients, and providers involved in a MaaS ecosystem grows, maintaining mutual trust relationships between all involved parties and orchestrating MaaS operations in centralized fashion quickly becomes intractable. We present a blockchain-based approach to providing MaaS in industrial settings where rented machines are equipped with IoT sensors, and where MaaS operations are orchestrated in a transparent, decentralized, and scalable way using a collection of smart contracts deployed over an infrastructure combining the Ethereum and InterPlanetary File System decentralized services. We detail the operations of MaaS, such as the lifecycle of management operations, and report on the performance of a prototype implementation deployed in the cloud.

Recommended citation: Tran, V. H., Lenssens, B., Kassab, A., Laks, A., Rivière, E., Rosinosky, G., & Sadre, R. (2022). Machine-as-a-Service: Blockchain-based management and maintenance of industrial appliances. Engineering Reports, e12567. http://hdl.handle.net/2078.1/265574

PProx: efficient privacy for recommendation-as-a-service

Published in Middleware, 2021

Presentation video Multi-tenancy enables cost-effective SaaS through resource consolidation. Multiple customers, or tenants, are served by a single application instance, and isolation is enforced at the application level. Ser- vice load for different tenants can vary over time, requiring applications to scale in and out. A large class of SaaS providers operates legacy applications structured around a relational (SQL) database. These applications achieve tenant isolation through dedicated fields in their relational schema and are not designed to support scaling operations. We present a novel solution for scaling in or out such applications through the migration of a tenant’s data to new application and database instances. Our solution requires no change to the application and incurs no service downtime for non-migrated tenants. It leverages external tables and foreign data wrappers, as supported by major relational databases. We evaluate the approach using two multi-tenant applications: Iomad, an ex- tension of the Moodle Learning Management System, and Camunda, a business process management platform. Our results show the usability of the method, minimally impacting performance for other tenants during migration and leading to increased service capacity after migration.

Recommended citation: Rosinosky, G., Da Silva, S., Ben Mokhtar, S., Négru, D., Réveillère, L., & Rivière, E. (2021, November). PProx: efficient privacy for recommendation-as-a-service. In Proceedings of the 22nd International Middleware Conference (pp. 14-26). http://hdl.handle.net/2078.1/252558

Active replication for latency-sensitive stream processing in Apache Flink

Published in IEEE SRDS, 2021

Presentation video Stream processing frameworks allow processing massive amounts of data shortly after it is produced, and enable a fast reaction to events in scenarios such as data center monitoring, smart transportation, or telecommunication networks. Many scenarios depend on the fast and reliable processing of incoming data, requiring low end-to-end latencies from the ingest of a new event to the corresponding output. The occurrence of faults jeopardizes these guarantees: Currently- leading high-availability solutions for stream processing such as Spark Streaming or Apache Flink’s implement passive replication through snapshotting, requiring a stop-the-world operation to recover from a failure. Active replication, while incurring higher deployment costs, can overcome these limitations and allow to mask the impact of faults and match stringent end-to-end latency requirements. We present the design, implementation, and evaluation of active replication in the popular Apache Flink platform. Our study explores two alternative designs, a leader-based approach leveraging external services (Kafka and ZooKeeper) and a leaderless implementation leveraging a novel deterministic merging algorithm. Our evaluation using a series of microbenchmarks and a SaaS cloud monitoring scenario on a 37-server cluster show that the actively-replicated Flink can fully mask the impact of faults on end-to-end latency.

Recommended citation: Rosinosky, G., Schmidt, F., Bodunov, O., Fetzer, C., Martin, A., & Riviere, E. (2021, September). Active replication for latency-sensitive stream processing in Apache Flink. In 2021 40th International Symposium on Reliable Distributed Systems (SRDS) (pp. 56-66). IEEE. http://hdl.handle.net/2078.1/252561

A Methodology for Tenant Migration in Legacy Shared-Table Multi-tenant Applications

Published in DAIS, 2021

Recommended citation: Rosinosky, G., Youcef, S., Charoy, F., & Rivière, E. (2021, June). A methodology for tenant migration in legacy shared-table multi-tenant applications. In Distributed Applications and Interoperable Systems: 21st IFIP WG 6.1 International Conference, DAIS 2021, Held as Part of the 16th International Federated Conference on Distributed Computing Techniques, DisCoTec 2021, Valletta, Malta, June 14–18, 2021, Proceedings (pp. 3-20). Cham: Springer International Publishing. http://hdl.handle.net/2078.1/250965

Élasticité de l’exécution des processus métier

Published in PhD thesis, 2019

The availability of middleware platforms in the cloud, with “transparent” scalability, is a progress for software developers and integrators. They can develop and deploy their applications without worrying about technical details. However, the exploitation cost of a cloud infrastructure can quickly become important. Providers requires methods to reduce this cost by adapting the size of ressources to the needs of the customers. In this thesis, we focus on multi-tenant transactional web applications, more precisely on business processes execution engines. We propose methods allowing to optimize the operational costs of providers of business process execution “as a Service” (BPMaaS) while ensuring a sufficient level of quality of service. This type of application do not scale well because of its persistence tier and of the transactional nature of operations. One must distribute the customers installations in order to optimize the cost, and sometimes move them depending of the needs of the customers. These moves (or migrations) have an impact on the quality of service and they must be limited. First, we propose a method for measuring the size of resources in terms of BPM tasks throughput, and then a method for measuring the impact of migrations we evaluate, thus confirming our hypothesis. We also propose several linear optimization models and heuristics targeting resouce allocation and distribution of customers, while limiting the number of migrations. These models are based on the knowledge of the needs of customers per time slot. We have experimented our three methods on the BPM solution Bonita, and demonstrated that they provide substantial savings on the infrastructure exploitation compared to a basic method.

Recommended citation: Rosinosky, G. (2019). Élasticité de l’exécution des processus métier (Doctoral dissertation, Université de Lorraine). https://hal.univ-lorraine.fr/tel-02096324

Business process outsourcing to the cloud: Balancing costs with security risks

Published in Computers in industry, 2019

The cloud computing introduced security risks which make companies reluctant to adopt it. Some means exist for reducing these risks while deploying a business process (BP) in the cloud. Among them, a technique largely used is to split the BP model in a choreography of BP fragments so that a cloud has only a partial view of the model. Unfortunately, reducing security risks with this approach implies the increase of BP fragments and of hosting clouds, and corollary the increase of costs. To overcome this issue, we present two heuristics managing separation/co-location constraints for risk versus cost optimization: the first optimizes the risk while respecting a cost threshold, the second optimizes the cost while respecting a risk threshold. These heuristics are experimented with a risk model, whose characteristics is to use only client-side data, and a more traditional cost model. Finally, the heuristics are validated against an exact optimization solution.

Recommended citation: Nacer, A. A., Godart, C., Rosinosky, G., Tari, A., & Youcef, S. (2019). Business process outsourcing to the cloud: Balancing costs with security risks. Computers in industry, 104, 59-74. http://hdl.handle.net/2078/235259

A Genetic Algorithm for Cost-Aware Business Processes Execution in the Cloud

Published in ICSOC, 2018

With the generalization of the Cloud, software providers can distribute their software as a service without investing in large infrastructure. However, without an effective resource allocation method, their operation cost can grow quickly, hindering the profitability of the service. This is the case for BPM as a Service providers that want to handle hundreds of customers with a given quality of service. Since there are variations in the needed load and in the number of users of the service, the allocation and scheduling methods must be able to adjust the cloud resource quantity and size, and the distribution of customers on these resources. In this paper, we present a cost optimization model and an heuristic based on genetic algorithms to adjust resource allocation to the needs of a set of customers with varying BPM task throughput. Experimentations using realistic customer loads and cloud resources capacities show the gain of these methods compared to previous approaches. Results show that using our algorithm on split groups of customers can provide even better results.

Recommended citation: Rosinosky, G., Youcef, S., & Charoy, F. (2018, November). A genetic algorithm for cost-aware business processes execution in the cloud. In Service-Oriented Computing: 16th International Conference, ICSOC 2018, Hangzhou, China, November 12-15, 2018, Proceedings 16 (pp. 198-212). Springer International Publishing. http://hdl.handle.net/2078/235253

Evaluating Multi-tenant Live Migrations Effects on Performance

Published in Coopis, 2018

Multitenancy is an important feature for all Everything asa Service providers like Business Process Management as a Service. Itallows to reduce the cost of the infrastructure since multiple tenantsshare the same service instances. However, tenants have dynamic work-loads. The resource they share may not be sufficient at some point intime. It may require Cloud resource (re-)configurations to ensure a givenQuality of Service. Tenants should be migrated without stopping the ser-vice from a configuration to another to meet their needs while minimiz-ing operational costs on the provider side. Live migrations reveal manychallenges: service interruption must be minimized and the impact onco-tenants should be minimal. In this paper, we investigate live tenantsmigrations duration and its effects on the migrated tenants as well asthe co-located ones. To do so, we propose a generic approach to measurethese effects for multi-tenant Software as a Service. Further, we proposea testing framework to simulate workloads, and observe the impact oflive migrations on Business Process Management Systems. The exper-imental results highlight the efficiency of our approach and show thatmigration time depends on the size of data that have to be transferredand that the effects on co-located tenants should not be neglected.

Recommended citation: Rosinosky, G., Labba, C., Ferme, V., Youcef, S., Charoy, F., & Pautasso, C. (2018, October). Evaluating multi-tenant live migrations effects on performance. In On the Move to Meaningful Internet Systems. OTM 2018 Conferences: Confederated International Conferences: CoopIS, C&TC, and ODBASE 2018, Valletta, Malta, October 22-26, 2018, Proceedings, Part I (pp. 61-77). Springer International Publishing. http://hdl.handle.net/2078/235247

Efficient Migration-Aware Algorithms for Elastic BPMaaS

Published in BPM, 2017

As for all kind of software, customers expect to find business process execution provided as a service (BPMaaS). They expect it to be provided at the best cost with guaranteed SLA. From the BPMaaS provider point of view it can be done thanks to the provision of an elastic cloud infrastructure. Providers still have to provide the service at the lowest possible cost while meeting customers expectation. We propose a customer-centric service model that link the BP execution requirement to cloud resources, and that optimize the deployment of customer’s (or tenants) processes in the cloud to adjust constantly the provision to the needs. However, migrations between cloud configurations can be costly in terms of quality of service and a provider should reduce the number of migrations. We propose a model for BPMaaS cost optimization that take into account a maximum number of migrations for each tenants. We designed a heuristic algorithm and experimented using various customer load configurations based on customer data, and on an actual estimation of the capacity of cloud resources.

Recommended citation: Rosinosky, G., Youcef, S., & Charoy, F. (2017, September). Efficient migration-aware algorithms for elastic BPMaaS. In Business Process Management: 15th International Conference, BPM 2017, Barcelona, Spain, September 10–15, 2017, Proceedings 15 (pp. 147-163). Springer International Publishing. http://hdl.handle.net/2078/235249

A Framework for BPMS Performance and Cost Evaluation on the Cloud

Published in IEEE CloudCom, 2016

In this paper, we describe a framework that allowsto automate and repeat business process execution on differentcloud configurations. We present how and why the differentcomponents of the experimentation pipeline -like Ansible,Docker and Jenkins- have been set up, and the kind of resultswe obtained on a large set of configurations from the AWSpublic cloud. It allows us to calculate actual prices regardingthe cost of process execution, in order to compare not onlypure performance but also the economic dimension of processexecution.

Recommended citation: Rosinosky, G., Youcef, S., & Charoy, F. (2016, December). A framework for BPMS performance and cost evaluation on the cloud. In 2016 IEEE International Conference on Cloud Computing Technology and Science (CloudCom) (pp. 653-658). IEEE. http://hdl.handle.net/2078/235245

An efficient approach for multi-tenant elastic business processes management in cloud computing environment

Published in IEEE Cloud, 2016

Even though the cloud computing paradigm has proven benefits, it faces a serious problem that can compromise its commercial success. It concerns the lack of efficient approach for using optimally the available resources. For this, several approaches have been proposed. However, they suffer from several shortcomings. Often only one objective is taken into account, expressing all operations in terms of cost. Furthermore, business processes should be insured with elasticity and multi-tenancy mechanism while adjusting the available resources to the dynamic load distribution. The proposed approach aims to optimize two conflicting objectives, namely the number of migrations of tenants and the cost incurred using a set of resources. It allows to take into account the multi-tenancy property and the Cloud computing elasticity, and is efficient as shown by an extensive experimentation based on real data from Bonita BPM customers.

Recommended citation: Rosinosky, G., Youcef, S., & Charoy, F. (2016, June). An efficient approach for multi-tenant elastic business processes management in cloud computing environment. In 2016 IEEE 9th International Conference on Cloud Computing (CLOUD) (pp. 311-318). IEEE. http://hdl.handle.net/2078/235234

Nouvelles fonctionnalités, évolutions et optimisations du logiciel d’assimilation variationnelle Yao

Published in Master thesis, 2012

The Yao software is a source code generator which aims to perform variational assimilation runs. One of its fonctionnalities is to help the user to conceive numeric simulations by using the notion of a modular graph for model representation. It gives an integrated development environment which provides the user with the ability to make standard and incremental variational assimilation sessions. Furthermore, it supplies a number of test and control functions. In this dissertation, we present the evolutions developed in the Yao software during the internship of Guillaume Rosinosky. A refactoring of the source code of the generated project has been done in order to simplify the maintenance and the readability of the source code, by using oriented object representations of the internal concepts of Yao. Two new algorithms have been incorporated and tested. The implementation of the quasi-static algorithm open the possibility to assimilate more complex models. The variational assimilation in weak constraint algorithm allows to take into consideration the model errors by switching in dual form the variational assimilation equations. In addition to those major modifications, general optimisation of the performances have been applied.

Recommended citation: Rosinosky, G. (2012). Nouvelles fonctionnalités, évolutions et optimisations du logiciel d’assimilation variationnelle Yao (Master dissertation, Laboratoire d'océanographie et du climat: expérimentations et approches numériques [LOCEAN], Institut Pierre Simon Laplace, Boîte 100, 4 place Jussieu 75252 Paris Cedex 05). https://dumas.ccsd.cnrs.fr/dumas-00985237