Mesos: A State-of-the-art Container Orchestrator

© 2016 Mesosphere, Inc. All Rights Reserved. 1
MESOS
A State-Of-The-Art
Container Orchestrator

InfoQ.com: News & Community Site
• 750,000 unique visitors/month
• Published in 4 languages (English, Chinese, Japanese and Brazilian
Portuguese)
• Post content from our QCon conferences
• News 15-20 / week
• Articles 3-4 / week
• Presentations (videos) 12-15 / week
• Interviews 2-3 / week
• Books 1 / month
Watch the video with slide
synchronization on InfoQ.com!
https://www.infoq.com/presentations/
mesos-api

Purpose of QCon
- to empower software development by facilitating the spread of
knowledge and innovation
Strategy
- practitioner-driven conference designed for YOU: influencers of
change and innovation in your teams
- speakers and topics driving the evolution and innovation
- connecting and catalyzing the influencers and innovators
Highlights
- attended by more than 12,000 delegates since 2007
- held in 9 cities worldwide
Presented at QCon San Francisco
www.qconsf.com

About me
Jie Yu (@jie_yu)
● Tech Lead at Mesosphere
● Mesos PMC member and committer
● Formerly worked at Twitter
● PhD from University of Michigan
● Worked on Mesos since 2012
http://people.apache.org/~jieyu/

● Mesos overview and fundamentals
● Why should I pick Mesos?
● Containerization in Mesos
Outline

● What does a traditional OS kernel provide?
○ Resource management Host cpu, memory, etc.
○ Programming abstractions POSIX API: processes, threads, etc.
○ Security and isolation Virtual memory, user, etc.
● Mesos: A kernel for data center applications
○ Resource management Cluster cpu, memory, etc.
○ Programming abstractions Mesos API: Task, Resource, etc.
○ Security and isolation Containerization
Mesos: A kernel for data center applications
Mesos overview and fundamentals

● Key concepts
○ Framework
○ Resource/Offer
○ Task
○ Executor
Programming abstractions
Master
Agent
Framework
Executor
Task Task
Executor
Task
Offer (Resources) Task/Executor
Resources Task/Executor

Case study: Marathon
Master
Agent X
Marathon
Offer
X: 8 cpus, 16G mem
Decline Offer
8 cpus, 16G mem

Create a Marathon app
Master
Agent X
Marathon
Executor
Task
Offer
X: 8 cpus, 16G mem
Accept Offer
LAUNCH(Task: 2 cpus, 2G mem)
POST /v2/apps

Create a Marathon app
Master
Agent X
Marathon
Executor
Task
TASK_RUNNING
TASK_RUNNING Offer
X: 6 cpus, 14G mem

A typical Mesos cluster
Master
Agent
Marathon
Agent Agent Agent Agent Agent Agent Agent
Kafka Cassandra MarathonSpark
Master Master
Zookeeper

Mesos helps improve cluster utilization
time
time

DS/OS vs. Mesos
Existing
Infrastructure
Mesosphere
DCOS
Services &
Containers
● Kernel alone is not enough
● DC/OS: the easiest way to run Mesos
○ CLI/UI
○ Package management
○ Service discovery
○ Load balancing
○ Day2 ops
○ Security
○ Framework SDK
● Yes, it is open source!

● Production ready
● Proven scalability
● Highly customizable and extensible
Why should I pick Mesos?
Why Mesos?

Production
Ready

The birth of Mesos
Why Mesos?
TWITTER TECH TALK
The grad students working on Mesos
give a tech talk at Twitter.
March 2010
APACHE INCUBATION
Mesos enters the Apache Incubator.
Spring 2009
CS262B
Ben Hindman, Andy Konwinski and
Matei Zaharia create “Nexus” as their
CS262B class project.
MESOS PUBLISHED
Mesos: A Platform for Fine-Grained
Resource Sharing in the Data Center is
published as a technical report.
September 2010
December 2010

Widely adopted
Why Mesos?
MESOS GRADUATES
Mesos graduates from the Apache
Incubator to become a top level
project.
June 2013
VERIZON SCALE DEMO
Verizon demonstrates launching
50,000 containers in less than 90
seconds using Mesos and
Mesosphere’s Marathon scheduler.
April 2013
MESOSPHERE
Mesosphere is formed by engineers
who have been using Mesos at
Twitter and AirBnB.
APPLE ANNOUNCES J.A.R.V.I.S.
Apple announces that the Siri
infrastructure now runs on Mesos,
atop “thousands” of nodes.
April 2015
August 2015

Production Mesos users
Why Mesos?

Proven
Scalability

Twitter
● Largest Mesos cluster
○ > 30000 nodes
○ > 250K containers

Apple
● Siri is powered by
Mesos!

Verizon
● 50K containers
in 50 seconds

● Stateless master
○ Inspired from the GFS design
○ Agents hold truth about running tasks (distributed)
○ Master state can be reconstructed when agents register
● Simple, only cares about
○ Resource allocation and isolation
○ Task management
● Implemented in C++
○ Native performance
○ No GC issue
Why Mesos is so scalable?
Why Mesos?

● Known that Mesos will scale to Twitter/Apple level
○ Feature is easy to add, took time to make it scalable
● Quality assurance for free
○ Imagine a test environment having 30k+ nodes with real workload
● Take backwards compatibility seriously
○ We don’t want to break their production environment
What does it mean to you?
Why Mesos?

Highly
Customizable
and Extensible

● Every company’s environment is different
○ Scheduling
○ Service discovery
○ Container image format
○ Networking
○ Storage
○ Special hardware/accelerators (e.g., GPU, FPGA)
● No one-fits-all solution typically
Why this is important?
Why Mesos?

Pluggable schedulers
Why Mesos?
● For instance, you need separate schedulers for
○ Long running stateless services
○ Cron jobs
○ Stateful services (e.g., database, DFS)
○ Batch jobs (e.g., map-reduce)
● Monolithic scheduler?
Monolithic schedulers do not make it easy to add new policies and specialized
implementations, and may not scale up to the cluster sizes we are planning for.
--- From Google Omega Paper (EuroSys’13)
Mesos frameworks
== pluggable schedulers

Flexible service discovery
Why Mesos?
● Mesos is not opinionated about service discovery
○ DNS based
○ ZK/Etcd/Chubby based (e.g., twitter, google, with client libraries)
○ Your custom way, every company is different
○ Mesos provides an endpoint to stream SD information
● DNS based solution does not scale well
Larger jobs create worse problems, and several jobs many be running
at once. The variability in our DNS load had been a serious problem for
Google before Chubby was introduced.
--- From Google Chubby paper (OSDI’06)

● Container image format
● Networking
● Storage
● Custom isolation
● Container lifecycle hooks
Pluggable and extensible containerization
Why Mesos?

● Mesos overview and fundamentals
● Why should I pick Mesos?
○ Pluggable architecture
○ Container image
○ Container network
○ Container storage
○ Customization and extensions
○ Nesting container support
Outline

What is Containerizer?
Containerization in Mesos
29
Containerizer
● Between agents and containers
● Launch/update/destroy containers
● Provide isolations between containers
● Report container stats and status
Mesos Master Mesos Master Mesos Master
Zookeeper
Marathon
Framework
Cassandra
Framework
Mesos Agent
Containerizer
Container
Executor
T1 T2
Mesos Agent
Containerizer
Container
Executor
T1 T2
Mesos Agent
Containerizer
Container
Executor
T1 T2

Docker containerizer
● Delegate to Docker daemon
Mesos containerizer
● Using standard OS features (e.g.,
cgroups, namespaces)
● Pluggable architecture allowing
customization and extension
Currently supported containerizers
Very stable. Used in large
scale production clusters

Mesos containerizer
● Support Docker, Appc, OCI (soon)
images natively w/o dependency

Unified containerizer
● Support Docker, Appc, OCI (soon)
images natively w/o dependency

● Pluggable architecture
● Container image
● Container network
● Container storage
● Customization and extensions
● Nesting container support
Unified Containerizer

Pluggable architecture
Launcher Isolators
Unified containerizer
Provisioner
Process
management
Container
lifecycle hook
Container
image support

35
Responsible for process management
● Spawn containers
● Kill and wait containers
Supported launchers:
● Posix launcher
● Linux launcher
● Windows launcher
Launcher

36
Interface for extensions during the life cycle of a container
● Pre-launch - prepare()
● Post-launch (both in parent and child context) - isolate()
● Termination - cleanup()
● Resources update - update()
● Resources limitation reached - watch()
● Agent restart and recovery - recover()
● Stats and status pulling - usage()
Isolator
Sufficient for most of
the extensions!

37
Isolator example: cgroups memory isolator
Agent Process
Launcher
creates
Subprocess Container
Process
execve()
LaunchInfo = Isolator::prepare()
* Create a cgroup for the container
in memory cgroup hierarchy:
/sys/fs/cgroup/memory/mesos/…
* Start listening for OOM event
Isolator::isolate(pid)
Block on pipe
Move ‘pid’ to the
memory cgroup just
created
Invoke ‘LaunchInfo.script’
Exec the executor
Signal the Child to continue

38
Agent Process
Container
Process
Isolator::update()
Change cgroup control:
memory.limit_in_bytes
Sending a new Task to
Executor, ‘resources’ of
the Executor changes
Send Task to Executor

39
Agent Process
Container
Process
Isolator::cleanup()
Remove the memory
cgroup associated
with the container
Shutdown Executor
or kill Task Destroy container
Container terminated

40
Cgroups isolators: cgroups/cpu, cgroups/mem, ...
Disk isolators: disk/du, disk/xfs
Filesystem isolators: filesystem/posix, filesystem/linux
Volume isolators: docker/volume
Network isolators: network/cni, network/port_mapping
GPU isolators: gpu/nvidia
…... and more! Need your contribution!
Built-in isolators

41
Start from 0.28, you can run your Docker container on
Mesos without a Docker daemon installed!
● One less dependency in your stack
● Agent restart handled gracefully, task not affected
● Compose well with all existing isolators
● Easier to add extensions
Container image support

42
● Mesos supports multiple container image format
○ Docker (without docker daemon)
○ Appc (without rkt)
○ OCI (ready soon)
○ CVMFS (experimental)
○ Host filesystem with tars/jars
○ Your own image format!
Pluggable container image format
Used in large scale
production clusters

43
● Manage container images
○ Store: fetch and cache image layers
○ Backend: assemble rootfs from image layers
■ E.g., copy, overlayfs, bind, aufs
● Store can be extended
○ Currently supported: Docker, Appc
○ Plan to support: OCI (ongoing), CVMFS
○ Custom fetching (e.g., p2p)
Provisioner

45
● Support Container Network Interface (CNI) from 1.0
○ A spec for container networking
○ Supported by most network vendors
● Implemented as an isolator
○ --isolation=network/cni,...
Container network support

46
● Proposed by CoreOS :
https://github.com/containernetworking/cni
● Simple contract between container
runtime and CNI plugin defined in the
form of a JSON schema
○ CLI interface
○ ADD: attach to network
○ DEL: detach from network
Container Network Interface (CNI)
Mesos Agent
Containerizer
Container
Executor
T1 T2
CNI Plugin
IPAM
veth
Network

● Simpler and less dependencies than Docker CNM
● Backed by Kubernetes community as well
● Rich plugins from network vendors
● Clear separation between container and network management
● IPAM has its own pluggable interface
47
Why CNI?

48
Existing CNI plugins
● ipvlan
● macvlan
● bridge
● flannel
● calico
● contiv
● contrail
● weave
● …
CNI plugins
You can write your own plugin,
and Mesos supports it!

49
● Support Docker volume plugins from 1.0
○ Define the interface between container runtime and storage provider
○ https://docs.docker.com/engine/extend/plugins_volume/
● A variety of Docker volume plugins
○ Ceph
○ Convoy
○ Flocker
○ Glusterfs
○ Rexray
Container storage support

50
Launcher
● Custom container processes management
Isolator
● Extension to the life cycle of a container
Provisioner
● New type of images
● Custom fetching and caching
Extensions

● New in Mesos 1.1
○ Building block for supporting Pod like feature
● Highlighted features
○ Support arbitrary levels of nesting
○ Re-use all existing isolators
○ Allow dynamically creation of nested containers
Nested container support

Mesos Master Mesos Master Mesos Master
Zookeeper
Marathon
Framework
Cassandra
Framework
Mesos Agent
Containerizer
Container
Executor
T1 T2
Mesos Agent
Containerizer
Container
Executor
T1 T2
Mesos Agent
Containerizer
Container
Executor
T1 T2
Container
Executor
T1 T2
Nested Container Nested Container

New Agent API for Nested Containers
message agent::Call {
enum Type {
// Calls for managing nested containers
// under an executor's container.
LAUNCH_NESTED_CONTAINER = 14;
WAIT_NESTED_CONTAINER = 15;
KILL_NESTED_CONTAINER = 16;
}
}

Launch nested container
Container
Executor
Mesos Agent
Containerizer
LAUNCH
Nginx

Watch nested container
Container
Executor
Mesos Agent
Containerizer
WAIT
NginxExit Status = 0

Arbitrary levels of nesting
Container
Executor
Nginx
Mesos Agent
Containerizer
LAUNCH
Debug

© 2016 Mesosphere, Inc. All Rights Reserved.
● Mesos: state of the art container orchestrator
○ Production ready
○ Proven scalability
○ Highly customizable and extensible
○ Pluggable architecture
○ Native support for Docker/Appc images (w/o Docker daemon or rkt)
○ Container network: CNI
○ Container storage: DVD
○ Nested container support
58
Summary

Questions?

60
CNI support using an isolator
Agent Process
Launcher
creates
Subprocess Container
Process
execve()
LaunchInfo = Isolator::prepare()
Tell the launcher to create the
child process in a new NET, UTS
and MNT namespace.
Isolator::isolate(pid)
Block on pipe
Bind mount the NET
namespace to keep
it open
Invoke ‘ADD’ of the
CNI plugin with the
NET namespace
associated with pid
Setup network
related /etc/xx files
for the container
Invoke ‘LaunchInfo.script’
Exec the executor
Signal the Child to continue

61
CNI support using an isolator
Container
Process
Isolator::cleanup()
Invoke ‘DEL’ of the
CNI plugin with the
NET namespace
handle
Umount and remove
the NET namespace
handle
Shutdown Executor
or kill Task Destroy container
Container terminated
Agent Process

Mesos, as one of the most powerful container orchestrators, greatly simplifies the
deploy, provision and execution of containerized workloads. It automates the
distribution of preprovisioned container images, injection of configuration,
scheduling onto machines, life-cycle-management, and monitoring of applications,
microservices, and jobs in the cloud.
In this talk, Jie Yu will first give you an overview about Mesos and its powerful API
which allows users to easily deploy their stateless and stateful services. Then, Jie will
talk about how containers are managed in Mesos. In particular, Jie will provide a deep
dive into the unified containerizer which is first introduced in Mesos 1.0.
Jie will show some of the new container networking and storage features that are
built recently, and how they benefit from the pluggable and extensible architecture of
the unified containerizer. Finally, Jie will discuss the future of container support in
Mesos.

Watch the video with slide
synchronization on InfoQ.com!
https://www.infoq.com/presentations/
mesos-api

Mesos: A State-of-the-art Container Orchestrator

More Related Content

What's hot (20)

Viewers also liked (16)

Similar to Mesos: A State-of-the-art Container Orchestrator (20)

More from C4Media (20)

Recently uploaded (20)

Mesos: A State-of-the-art Container Orchestrator