#MesosCon 2015

As a distributed systems kernel, Mesos solves complex scheduling problems. These problems include: Bin Packing of tasks given resource constraints; Dominant Resource Allocation to prioritize Framework offers; user-defined optimistic or pessimistic Scheduler algorithms. When job allocation stops working properly, we see Framework starvation, dead locks, underutilization and other issues. In this talk, Alex Gaudio will present a unique interpretation of task scheduling on Mesos. By drawing on his experiences building Mesos Frameworks and infrastructure for Sailthru's Data Science team, he will provide attendees with ways to fundamentally question how they use Mesos, and offer practical advice to improve sub-optimal usage patterns.

The previous Mesos API provided language bindings (a.k.a. drivers) for frameworks to interact with Mesos. Framework developers don’t like dealing with dependencies in native libraries, and with Mesos’ new HTTP API, they no longer have to! In this session, (Vinod Kone, Technical Lead for Mesos at Twitter and Isabel Jimenez, Software Engineer at Mesosphere) will provide a comprehensive walk-through of recent advancements with the Mesos API, explaining the design rationale and highlighting specific improvements that simplify writing frameworks to Mesos. 

Netflix customers stream more than two billion hours of content each month, accounting for over a third of downstream Internet traffic. At this scale, real time insights are critical to a smooth streaming experience. Sharma Podila will present an in-depth look at the technology - including the Fenzo task scheduler for Apache Mesos and Mantis, a cloud native event stream processing platform - that Netflix uses to gain real time operational insights into system behavior. The talk will focus on the design and usability of Fenzo and will provide details on how it uses CPU, memory, and network bandwidth resources for job scheduling. The presentation will showcase bin packing, task locality, resource affinity, task constraints, and autoscaling of the slave cluster. To close Sharma will talk about future directions and opportunities to collaborate with the Mesos community.

Kubernetes provides a declarative model in order to describe micro-services. However, auto-discovering those services outside of kubernetes requires add-ons. In this presentation we will explore how using the kubernetes framework and Mesos-DNS we can bridge the gap, and provide cluster wide service discovery.

One of the most exciting new innovations is a simple, pluggable approach to building customizations for Mesos. This makes it as easy to extend or customize Mesos as it is to install plug-ins into your web browser. In this talk, (Niklas Nielsen and Kapil Arya, Distributed Systems Engineers at Mesosphere) will provide a technical introduction to Mesos Modules - the new way to customize Mesos without having to recompile / relink for each specific use case. They will demonstrate how to build and use Mesos Modules, including an example of how to write an allocator module for custom scheduling policies across a cluster. The presenters will also explain how Mesos Modules will help grow the ecosystem of developers building new features for Mesos. Customization will be one of strongest and most flexible attributes of the Mesos distributed systems kernel!

Real Time Bidding on the large Internet exchanges (Doubleclick, Facebook, etc) requires large-scale, low-latency serving systems: >100K QPS at peak with <100ms tail response times. Time is very literally money. Dobromir will present TellApart's full stack in excruciating detail (if you want it), which includes Mesos/Aurora, ZK service discovery, Finagle-Mux RPC, and a Lambda architecture with Voldemort as the serving layer.