A few weeks ago, Sabri Skhiri and Florian Demesmaeker were in London to attend the Spark+AI summit. They came back with a lot to say about the new features of Spark and the presented use cases! In this article, they will give you their opinion about Databricks’ main announcement, the intakes of their favourite talks and training, and what they thought of the new name of the conference.
A new name
This year, Spark expanded the summit’s scope and renamed it “Spark + AI Summit”. The goal of Databricks, announced by its co-founder Ali Ghodsi, is to incorporate unified aspects of data and AI.
Florian Demesmaeker, our R&D engineer, explains: “In some of the keynote talks, the speakers talked about use cases where the job of the data engineer is strongly reduced. The data scientists can easily experiment with data, travelling back and forth in time. This means more focus on AI, rather than on the data engineering part that makes all data accessible to the data scientists”.
In line with this change of name, Databricks announced the release of a complete data science lifecycle on the cloud.
Sabri Skhiri, our R&D Director, explains “It is interesting to see that the change in the event name is actually very visible in the change of Databricks’ strategy. Their tools are now completely dedicated to stream ETL, and there is a huge focus on integrated data management”.
Databricks’ new features include Databricks Delta which creates data pipeline and provides data views and exploration features. Secondly, the Databricks Runtime ML is a ready-to-use environment providing a set of pre-loaded ML frameworks where the data scientist can play with data. Finally, the MLflow tool allows to simplify the ML models development at enterprise scale.
Our R&D Director precises: “Together, these features provide a complete and unified approach to machine learning lifecycle and pipeline automation. This looks like a very competitive SaaS offer for integrated data management, available on AWS and Azure. However, the metadata management and the security aspect is still the missing piece”.
The training day
The first day of the conference was dedicated to training workshops that include a mix of instruction and hands-on exercises to help attendants improve their Apache Spark skills.
Florian gives insights into his favourite training Tuning and Best Practices. He explains: “The aim of the training was to make programmers aware of how Spark works internally, in order to be able to write optimised applications. They presented a few situations, each one showing one relatively slow process. Then they presented a step-by-step procedure to debug the situation and to find the points that could be improved in the current situation. In summary, tips and tricks to adapt to different situations”.
The sessions at the conference covered data engineering and data science contents along with best practices for productionising AI. The talks were divided into roughly two categories: Spark programming and deployment, and applications on top of Spark (AI applications).
Florian Demesmaeker explains: “I attended 28 talks. The keynotes from Databricks were quite interesting, they presented Delta and MLflow. I also enjoyed the talks about tools to optimise the internals of Spark, these provided good technical details. Other talks were about use cases on top of Spark, it was interesting to see what challenges other companies face and how they address them”.
Sabri Skhiri adds: “The talk Learning to Rank Datasets for Search was very inspiring. Oscar Castañeda-Villagrán, a data scientist working at Xoom (a Paypal service) talked about learning to rank R data set. The idea is that we can extract metadata when the data pipeline is arriving in the lake. Going further, you can not only extract metadata but also calculate a kind of judgment relevance score that will be used for bootstrapping the learning to rank process. In this way, a user can search and retrieve the relevant R data set in the lake. A very good idea for the metadata-driven exploration”.
Early September 2018, 8 EURA NOVA engineers travelled to Berlin to attend the Flink Forward Conference, dedicated to Apache Flink users and stream processing communities. You can read their feedback here.
Our paper “Data Mining and Machine Learning Techniques supporting Time-based Separation Concept Deployment”, co-written with Eurocontrol and WaPT, has been accepted by the 37th Digital Avionics Systems Conference (DASC) in London, U.K.
The paper presents two methods to allow air traffic controllers to deliver separation minima accurately and safely, on the basis of time intervals instead of distances.
Importantly, in strong headwind conditions, the aircraft’s groundspeed during approach decreases, meaning that keeping the distance-based separation method results in lower landing rates. At a time of intensified air traffic, this situation leads to considerable delays at airports with significant costs to operators and travellers.
With the new methods presented in the paper, capacity can increase by up to 14% in strong wind conditions, and by up to 8% in moderate wind conditions.
The paper will be presented in September at DASC 2018, but you can already read the abstract below. If you wish to go deeper into the subject, do not hesitate to contact our research department at firstname.lastname@example.org.
The Time-Based Separation (TBS) concept consists in the definition of separation minima for aircraft on the final approach to a runway based on time intervals instead of distances, as applied in Distance-Based Separation (DBS) operations.
TBS allows for dynamic distance separation reductions in strong headwind conditions so as to preserve time spacing across all wind conditions. However, TBS application entails the use of a support tool providing separation distance indicators depending on the applicable time separation minimum, the aircraft speed profile which also depends on the headwind conditions.
This paper details two methodologies allowing a system to compute those TBS indicators so as to allow Air Traffic Controllers to accurately and safely deliver the TBS minima using a separation delivery support tool. The first approach is based on “analytical” data mining and modelling whereas the second one is based on a Machine Learning (M/L) procedure.
In the framework of the deployment of the TBS concept in Vienna airport (LOWW), those approaches are developed and tested using a database covering one year of traffic and corresponding local meteorological data.
The operation of TBS with indicators computed using either approaches leads to substantial diminution of time separations compared to a DBS strategy. However, given the large uncertainties related both to leader and follower aircraft speed profiles, the buffers could be designed only for the most frequent pairs. With the M/L approach (resp. the “analytical” approach), the capacity benefits related to the application of TBS with a separation support tool are of the order of 8% (resp. 2%) in moderate wind conditions, and up to 14% (resp. 10%) in strong wind conditions.
Iterative-convergent algorithms represent an im-portant family of applications in big data analytics. These aretypically run on distributed processing frameworks deployed on a cluster of machines. On the other hand, we are witnessing the move towards data center operating systems (OS), where resources are unified by a resource manager and processing frameworks coexist with each other. In this context, different processing framework job tasks can be scheduled on the same machine and slow down a worker (straggler problem). Existing work has shown that an iteration model with relaxed consistency such as the Stale Synchronous Parallel (SSP) model, while still guaranteeing convergence, is able to cope with stragglers. In this paper we propose a model for the integration of the SSP model on a pipelined distributed processing framework. We then apply SSP on a distributed version of the Frank-Wolfe algorithm. We theoretically show its sparsity bounds and convergence under SSP. Finally, we experimentally show that the Frank-Wolfe algorithm applied on LASSO regression under SSP is able to converge faster than its BSP counterpart, especially under load conditions similar to those encountered in a data center OS.
Nam-Luc Tran, Thomas Peel, Sabri Skhiri, Distributed Frank-Wolfe under Pipelined Stale Synchronous Parallelism, proceedings of the 2015 IEEE Conference on Big Data, November 2015, Santa Clara, CA, USA.