The Cloud Layer of the RDC is the technical backbone based on a multi-cloud infrastructure, notably consisting of the de.NBI cloud and GDWG. It provides near-infinite storage and storage services such as Aruna Object Storage (AOS), which manages the RDC's raw datasets. In addition, the layer provides basic compute services to the upper layers to facilitate applications to run in the cloud. The cloud layer is not designed for the end user but for those who implement and deliver the services on the upper layers.
Zooming into the Cloud Layer
The Cloud Layer provides support for storage and compute resources. For the storage resources, we distinguish between primary storage, with its reference implementation Aruna Object Storage, and secondary storage, where dedicated data systems are offered to manage the data products of the Semantic Layer. The compute resources consist of tools, such as Docker and Kubernetes, to facilitate the depolyment of services of the upper layers in the cloud.
Primary Storage - Infrastructure-as-a-Service
The primary storage consists of a cloud-based object storage, in our case the Aruna Object Storage (AOS), developed at Justus Liebig University Giessen (JLU). An object storage offers several advantages: As indicated by the name, an object storage organizes data in units of objects (not for example as files as known from a file system). Objects have a system-generated unique identifier and metadata to describe the contents of the objects. The object creator is responsible to attach suitable metadata that describe the contents of unstructured data data like text, images, audio and video or semi-structured data expressed in JSON format and support finding objects. In order to retrieve objects from the object storage, there are two options. One is to use the unique identifier to retrieve the associated object and the other makes use of search labels, attached to the objects as metadata, returning all objects matching one or multiple search labels. Second, one of the key functionality of an object storage is the possibility to use so-called RESTful APIs or HTTP to retrieve the objects from the storage. Thus, it is possible to use either an arbitrary programming language or simply a web browser access the data. While object stores might be located on an ordinary desktop computer, an object storage is distributed among nodes in the cloud. This gives at least three advantages: First, the cloud offers a near unlimited storage capacity. Second, it is possible to distribute objects among multiple nodes such that access to objects and processing can be parallelized resulting in large performance improvements. Third, the data objects are replicated to avoid the problem of system failures. When one of the nodes in the cloud is out of service, there will be other nodes offering a copy of the objects.
The object storage of a research infrastructure like RDC gives another advantage by sharing objects among many users. Instead of managing an individual copy of an object for each of the users, it is now possible to share the object among all users. Moreover, domain scientists do not have to care about the management of systems, but simply use the unified access interface of the object storage of RDC.
In general, object storage requires that objects are static and updates of an object are seldom. However, in biodiversity there are highly dynamic data sets such as time series that need to be updated continuously. To address this issue, RDC introduces a version concept that manage static versions, also known as snapshots, of a data set, where one of them is the current version. Instead of applying updates to the actual version immediately, it is possible to collect updates and apply them once when the next version of the data set is created. Versions of data sets are then managed for processing and never deleted from the object store. Thus, such a version approach also supports reproducibility of research results.
Cloud Computing - Platform as a Service
The RDC is not limited to storage only, but strives for offering scalable compute services that are easy to use based on intuitive low-code interfaces. We aim to offer these low-code services in the Mediation Layer and are currently implementing them using the popular and powerful open-source Apache Spark infrastructure as a PaaS (Platform as a Service) for data transformation related tasks. Building on such infrastructure lets us benefit from other existing developments on top of it so we can avoid building things from scratch and focus our efforts on specific challenges present in the RDC.
For enabling tools like Apache Spark to run in the cloud in a distributed and scalable fashion, we build on Kubernetes for managing compute resources available on cloud providers like de.NBI. In order to remove the complexity of deploying Apache Spark in a Kubernetes cluster for end-users, we want to offer Apache Kyuubi as a distributed JDBC/ODBC gateway service. This enables the submission of Spark tasks in SQL-like query syntax and Kyuubi handles the creation and management of Spark compute instances in the Kubernetes cluster. At the same time, it enables easy integration into other tools, since only a JDBC/ODBC client is required for interfacing with the service. For example, on the Semantic Layer, we offer a Jupyter Notebooks service, from which data transformations can then seamlessly be triggered to run on Spark, offering the power of the cloud to our users.
Perspectives
Secondary Storage - Platform as a Service
Unlike primary storage, secondary storage provides data systems as a service that are managed in the cloud. For example, users are often interested in using a relational database system such as PostgreSQL with its native SQL interface. For example, some tools use a PostgreSQL database to manage geospatial data for species observations or environmental parameters. Instead of deploying such a service locally, we strive to provide such services for users to create databases. In particular, we see the need for such a service for managing data products created in the Semantic Layer. These data products often come with a structured or semi-structured data model (e.g. JSON). It would then be advantageous to use an appropriate data service rather than hosting a database system on a local machine. This would also potentially open up data sharing among users and leverage collaborations within a community. In addition, cloud-ready systems are often available, and the scalability of the cloud would provide a performance advantage. We have not yet implemented such services, but we plan to support various data services such as PostgreSQL, NoSQL database systems, and triple stores, as these systems are already required within the RDC.
Associated Services: