Building the Glue for Service Discovery & Load Balancing Microservices

© 2016 Mesosphere, Inc. All Rights Reserved.
BUILDING THE GLUE FOR
SERVICE DISCOVERY &
LOAD BALANCING
MICROSERVICES
@SARGUN
1
Sargun Dhillon, 2016

WHO AM I?
2

AGENDA
3
• How, and why did we go here?
• What have I been building in the past year?
• Why?
• How?
• Challenges
• Some of the future

A BRIEF HISTORY
OF NETWORKS
IN THE
DATACENTER
4

ORGANIZATIONS CIRCA 2007
5
•Before DevOps was coined
•Clear diﬀerentiation of ownership
•The datacenter was owned by the NOC
•Deployment of services was done by sysadmins in the operations group
•Developers operated without access to production systems
•Production deployments gated by QA and Operations

SOFTWARE CIRCA 2007
6
•Diﬀerent services glued together via CORBA, XML-RPC, SOAP
•No one was really consciously doing microservices
•Networks were static, giant layer 2 domains
•Load Balancing provided by hardware
•Everyone ran their own datacenter
•EC2 in its infancy, only a year prior has the term “Cloud” began to become popular
•Systems statically partitioned

© 2016 Mesosphere, Inc. All Rights Reserved. 7
SaaS continued to grow at
an incredible rate

There became a race to ship faster

We kept the software alive
By feeding it
With Sysadmins

We kept the machines alive
By feeding them
With Blood

This wasn’t working

ORGANIZATIONS 2008+
12
•We began seeing a gradual shift in the industry where lines between QA, Dev, and
Ops were blurring
•Devops term coined in 2008, first DevOpsDay in 2009
•Gradual adoption of the cloud, fewer organizations owning their own datacenters
•Either networking was outsourced to the cloud, or typically remained in a small
internal organization
•Needed to reduce ratio of operators to servers

SOFTWARE CIRCA 2008+
13
•Popularization of Open Source tooling to automate QA and much of traditional operations
•Jenkins / Hudson
•Puppet / Chef
•Capistrano
•Popularization of stacks requiring more complex operational requirements
•Nutch / Hadoop
•NoSQLs
•Machine still statically partitioned
•Networks still sacred territory

CIRCA 2011
14
•Much of what’s been happening for the past half-decade hits networking
•Much of this falls under the term “SDN” (Software Defined Networking) or “NFV” (Network
Function Virtualization)
•Hastened by the adoption of VMs in the enterprise in the hype cycle
•Openflow promises to fix everything
•Major adoptions of the cloud by startups as well as enterprise
•Virtualization begins to become mainstream as a mechanism of consolidating workloads
•The invention of the “private cloud”
•DotCloud / Docker funded by Y-combinator a year earlier
•Term “Microservice” coined

CIRCA 2013
15
• Docker becomes an instant hit and brings containers to the forefront
• Dynamic partitioning begins to make in-roads
• Google releases Omega paper
• Apache Aurora open sourced
• Microservice counts explode, demanding collocation of workloads for eﬀiciency
• Mesosphere Founded
• Site Reliability Engineering begins to popularize and further blur the lines between
Dev, Ops, and QA

Everything was changing

Why?

Business Value

BENEFITS
20
•Reduction in cost of goods sold
•Smaller engineer to server ratio
•Linear, or super linear growth rate of engineering team to servers is unsustainable
•Smaller engineer to capability ratio, where capability includes:
•Features
•Throughput
•Better User Experience
•Better availability
•Quicker release to features and fixes

But at what cost?

Complexity

OLD WORLD
23

NEW WORLD
24

DIVING DEEPER
25

Paxos?
Raft?
Ω Failure Detector?
Pods?
Wat?
Sidecars?
Etcd?
Zookeeper?
VxLan?

Performance

REDIS PERFORMANCE
28

MySQL Performance with Containers
0
75000
150000
225000
300000
Container-free Host Mode Bridged Overlay
Transactions / Sec

OUR JOURNEY
30

“Connectivity”

Where are my apps running?

The Old World

Let Mesos* Choose Ports
*The Scheduler

How do you find the tasks?

A Directory?

See: DNS

Solution:
Expose the IP and Port of Tasks
via
DNS SRV records

Everyone has DNS right?

And GLibc even has a bug open for it!

…Opened in 2005

So, we performed an OODA loop
1.Observe
2. Orient
3. Decide
4. Act

OBSERVE: SERVICE DISCOVERY
45
• Existing Dynamic Service Discovery Solutions:
• Etcd
• Finagle + Zookeeper
• Consul
• Existing Static Service Discovery Solutions:
• Amazon ELB
• Hardware Load Balancers
• Service Discovery is an afterthought
Gathering our data about the field

OBSERVE: NETWORKING
46
• Everybody assumes IP per application instance
• Everybody assumes reliable DNS
• Some people want to be fast
• Some people want security
• Nobody wants to edit application code
• Nobody wants to talk to their network engineer
Gathering our data about the field

ORIENT:
NETWORKING
CORE TENANTS
47
•We must be agnostic to the underlying environment
•AWS / Azure / GCE / Softlayer as the lowest common
denominators
•We should require minimal changes to user code in
order to work
•We should provide similar services to existing
environments
•Fixed load balancers
•Security
•IP/Container
•We do not want to require a change in organization
procedures

Act:
What did we build?

Load Balancing:
Minuteman

WHAT WE WANTED
50

First try:
LD_PRELOAD

How does connect() work?

How does connect() work on
LD_PRELOAD?

…But no

Static Linking

What else is there?

Minuteman:
Flows at a Moment’s notice

MINUTEMAN:
BENEFITS
68
•Appearance of a fixed-load balancer
•Fully distributed
•Other than first packet, the entire lifetime is handled
in kernel space
•Source: github.com/dcos/minuteman

How do we tie it all together?

GLOBAL STATE
70
• Load Balancer Task Mapping
• DNS Zones
• Virtual Network Routing Tables
• Reachability
• Security ACLs

Computers

Constant Churn

Sometimes
They Break

Sometimes
Many Break

Sometimes
You don’t know

HOW DO YOU
DO
SIGNALING?
76

Nobody ever got fired for using
Zookeeper

We know about Zookeeper

It works…usually

How else can we do this?

Naively

Massive Amount of Information

Who else has solved this?

Academia

Control Plane:
Lashup

How do we scale the naive approach?

All we need is a sparse, connected graph

But how?

Enter: HyParView
Builds a connected graph (overlay), where the degree is <=K

Boot Time

Connected Graph
Hyparview

Constant Adaptive
Health Checks*
*Borrowed from SWIM, Gossip Style Failure Detector

Dealing with Failure
Hyparview

Hyparview

Routing*
*Borrowed from Dijkstra, OSPF, IS-IS, Perlman

Just Run Dijkstra*
*BFS / DFS

Minimum Spanning Tree, From Sender A

A Wild Pokemon Appears!*
*CRDT papers
Database

CRDTS: Semilattices
“Time”

CRDT KV Store

LASHUP KV
118
•Datatypes:
•Maps
•Composable
•Sets
•Flags
•Counters
•Registers
•Last-write-wins

Evaluation

Prior to Lashup

After Lashup

PROJECT
LASHUP
122
•A novel distributed systems SDK that provides:
•Membership
•Multicast Delivery
•Strongly-eventually consistent data storage
•Powers:
•Minuteman VIP dissemination
•Minuteman node liveness checks
•Overlay routing
•DNS Synchronization
•Source: github.com/dcos/lashup

THE FUTURE
123

Zero Overhead NFV:
Software Defined Systems?

This seems familiar

Something new

Something new(-ish)

(Extended)BPF?

ZERO-OVERHEAD NFV
132
•Implemented using built-in Linux APIs
•Linux Security Module API
•Linux eBPF
•Allows standard BSD API to work
•Example: getpeername() works
•In preliminary testing: <0.1% overhead
•Challenges
•Upgrading Kernel

IPv6:
Just more addresses?

000.000.000.000
32-bits
IPv4

0000:0000:0000:0000:0000:0000:0000:0000
64-bits
IPv6
64-bits
Network
Part
Host
Part

SLAAC
StateLess Address Auto Configuration

IPV6
137
•264 addresses per host
•Automatically configured
•Challenges
•Cloud support
•Organization challenges

What did we learn?

Organizations are weird

We’re probably doing it wrong
(today)

The future looks bright

BUILDING THE GLUE FOR
SERVICE DISCOVERY &
LOAD BALANCING
MICROSERVICES
@SARGUN
142

Building the Glue for Service Discovery & Load Balancing Microservices

Related slideshows

More Related Content

Building the Glue for Service Discovery & Load Balancing Microservices