Hacker News Reader

Comment by wmf | Hacker News Reader

Maybe this is a good place to ask a question that I've been pondering for a while: Why is Mesos based on the concept of resource offers? AFAIK this is backwards compared to other schedulers where you ask for resources and the schedule gives them to you (or not).

View on HN

I can't quite tell you why it was implemented like this. The Omega paper by Google describes different kinds of scheduling algorithms and their pros and cons.

I don't think that it's backwards compared to other schedulers (many of them are monolithic) but I think it can be improved by using transactions and optimistic concurrency control. This is discussed in the Omega paper and I think there's work to integrate the shared state scheduler into Mesos.

The most puzzling thing for me is why schedulers and executors are so tightly coupled in Mesos. In my mind, those are separate concerns: schedulers should concern themselves solely with locating an available resource for you (and should be part of the Mesos core), and the executor should take care of launching and managing the program to be run.

With the current design, if I want to launch a Docker container, I have to use its scheduler, even though its scheduler might not be as good as some other resource's scheduler. And I might have to accept a regression in the scheduler in order to benefit from a bugfix in the executor.

To load the framework (application) with the logic with the concern of deciding what resources it wants. The scheduler shouldn't care about what you get, just fairness. Two-tiered scheduling achieves this:

Mesos decides how many resources to offer each framework, while frameworks decide which resources to accept and which computations to run on them.

http://mesos.berkeley.edu/mesos_tech_report.pdf

I guess that makes sense if frameworks care about placement (which I generally don't) because otherwise they'd have to express placement constraints to Mesos.

I still can't find documentation about what fairness policies are actually implemented, though.

Mesos uses dominant resource fairness:

https://www.cs.berkeley.edu/~alig/papers/drf.pdf

Reply on news.ycombinator.com