Version 4.6.0.0 is a major release from the technological point of view (really big changes happened behind the hood) while in terms of functionalities, very little visible changes happened. In terms of invisible changes, two major milestones mark this release. First of all, the architecture running EPPI-Reviewer has been moved: EPPI-Reviewer now runs in a fully virtualised environment hosted on Microsoft Azure. Secondarily, the Machine Learning portions of EPPI-Reviewer (also known as Priority Screening) are now fully automated (previously some important preparation steps where semi-manual) and is now running entirely on the Azure Machine Learning environment.
The first change has some implications in terms of reliability, service levels and backup provision, the second will allow important improvements to the screening phase to become public as soon as the new features will be tested by a few early adopters. If you wish to participate in the beta testing phase, please do get in touch and we'll enable Priority Screening for you. The following will describe the most important changes and finish up with a short list of bug-fixes and minor improvements.
EPPI-Reviewer on Microsoft Azure.
Until Friday 22 April, the EPPI-Reviewer servers were hosted within the UCL Institute of Education. This meant our team had full control of all the configurations, both hardware and software, but created an important dependency: whenever a problem happened to the Internet connection of the University, EPPI-Reviewer could become inaccessible to both internal and external users. Unfortunately, problems did happen, the last one being caused by what appeared to be a sustained denial of service attack against the whole UK academic network, causing unplanned service interruptions during working hours. This dependency is now entirely removed, but of course, EPPI-Reviewer still requires an active Internet connection. However, in the new situation, connectivity is ensured for Microsoft data centres as a whole, so we are confident that reliability will be much higher.
Service levels and Data storage.
EPPI-Reviewer is now hosted in Azure Data Centres, in particular in what Microsoft calls "West Europe" zone, which means data-centres in the Netherlands. All data is kept there, with the exception of database backups which are replicated also into the "North Europe" zone (Ireland), for disaster recovery purposes. In terms of data retention, nothing has changed, our policy of not deleting user data (no matter how old) hasn't changed, users retain ownership of their data and we will not actively interfere. Legally, all data remains in the EU, so is subject to the same privacy and data protection legislation as before. What has changed, and will probably change again later this year, is the backup policy.
Backups Provision
Our previous backup system was Tape-based, and allowed us to keep tapes with full backups for long periods. This kind of arrangement is not possible on a fully virtualised environment, so we had to design an entirely new backup solution using Azure-based technologies. Without describing the technical details, the interim arrangement that we have established is as follows:
We will retain full backup data for 7 days. Within this time range, as long as no unexpected errors happen, we will be able to restore data from any arbitrary moment in time.
After 7 days we will not retain the full history of data changes, but only a snapshot of the full backup, one per day. These snapshots will be initially kept for 15 days. Within the next few weeks we will review the cost of such an arrangement, and if feasible we will extend the data retention period of daily snapshots.
Please also note that backup services are offered strictly on a "best effort" basis: it is notoriously difficult to be sure that all backup sets are functional without testing each one individually.
Regarding service availability, as mentioned above, EPPI-Reviewer is now entirely independent from the infrastructure of UCL Institute of Education, this will in general ensure a significantly improved level of service (understood as a decreased amount of unplanned downtime). However, every now and then (two or three times per year, based on historic data), Microsoft needs to apply changes to the underlying infrastructure. When this happens they will need to reboot our systems, and might keep them offline for up to ten minutes. Furthermore, Microsoft will inform us well in advance, but will provide a time window (usually during week-ends). As a result, we will be able to warn EPPI-Reviewer users about upcoming service interruptions well in advance, but we won't be able to specify the precise time when they will occur. This shouldn't be a major concern as EPPI-Reviewer saves data in real time so there is no particular risk of data loss.
Machine Learning and Priority Screening facilities
These are available on request, but we expect to make them open to all EPPI-Reviewer users very soon. The new system seamlessly integrates the machine learning side with a new and improved screening workflow, rich with additional automation options.
Minor changes and bugfixes:
Version 4.6.0.0 includes a new system to make the initial loading of reviews more reliable. Because of the architectural changes, the system used in previous version started showing some weaknesses, in particular it was prone to timing out, which can be quite disruptive. To fix this problem, the new system will try using a second strategy in case the first attempt fails, this has the secondary effect that loading up a large review (100,000 items or more) might take longer. On 09/05 we've added an additional improvement (with no detectable drawbacks) to further improve reliability as a few users experienced timeout problems on the same day.
Version 4.6.0.1 was a snap release to fix a problem in the Coding-Only interface: adding or removing a code would cause an unnecessary error message to appear, even if data was saved correctly; this problem is now fixed.