A Cloud-Based, Geospatial Linked Data Management System - Transactions on Large-Scale-Data-and Knowledge-Centered Systems XX - page 76

Fig. 2. The new architecture of the LD management system
over until the master instance is up again. As the master instance is a vital com-
ponent of the system, there is also a script in the instance hosting LMS which
is informed when the master instance is down and attempts to reboot it.
As it can be easily understood, the difference between a master and a slave
instance is that a master instance has triggers associated to the RDF QUAD
table of Virtuoso server and also includes the Updater component.
The image updating relies on the EBS-back up mechanism of Amazon which
enables to create snapshots of image volumes and then the respective AMIs out
of them which can then be used to create new instances in a rapid manner. This
also ensures that: (a) system data become reliable as they are stored across many
places in the Amazon Cloud and (b) as snapshots are incremental, the system
cost is increased only in terms of what is updated. When back-up takes place,
the master instance is shut down to ensure that image creation does not lead to
any file integrity issues. In this case, query and export requests are served by the
slave instances in the Amazon LB group while update requests are served by the
backup master instance. When the master instance is up again, it is updated by
the backup instance based on the RDF updating mechanism mentioned above.
The advantages of the proposed architecture are the following: (a) it is not
very costly: LMS can be hosted in a small-sized instance and new slave instances
are created only when the load is assessed to be heavy for the current instances
running while one or more created instances are deleted when the load is not so
heavy any more but after a particular period passed by such that we avoid any
circular scale-in and scale-out effect which can become quite costly, (b) updating
is eventually propagated to all instances running and to those instances that will
be created, (c) load balancing can guarantee a satisfactory performance level for
query/export requests which represent the majority of the user requests incurred,