Large-Scale Distributed Systems
Andrew Whitaker
CSE451
Textbook Definition
“A distributed system is a collection of loosely coupled processors interconnected by a communication network”
Typically, the nodes run software to create an application/service e.g., 1000s of Google nodes work together to
build a search engine
Why Not to Build a Distributed System (1)Must handle partial failures
System must stay up, even when individual components fail
Amazon.com
Why Not to Build a Distributed System (2) No global state
Machines can only communicate with messages
This makes it difficult to agree on anything “What time is it?” “Which happened first, A or B?”
Theory: consensus is slow and doesn’t work in the presence of failure So, we try to avoid needing to agree in the first place
A B
Reasons to Build a Distributed System (1)The application or service is inherently
distributed
Andrew Whitaker Joan Whitaker
Reason to Build a Distributed System (2)Application requirements
Must scale to millions of requests / sec Must be available despite component failures
This is why Amazon, Google, Ebay, etc. are all large distributed systems
Internet Service Requirements
Basic goal: build a site that satisfies every user requests
Detailed requirements: Handle billions of transactions per day Be available 24/7 Handle load spikes that are 10x normal capacity Do it with a random selection of mismatched hardware
An Overview of HotMail (Jim Gray) ~7,000 servers 100 backend stores with 300TB (cooked) Many data centers Links to
Internet Mail gateways Ad-rotator Passport
~ 5 B messages per day 350M mailboxes, 250M active ~1M new per day. New software every 3 months (small changes weekly).
Availability Strategy #1: Perfect Hardware
Pay extra $$$ for components that do not fail
People have tried this “fault tolerant computing”
This isn’t practical for Amazon / Google: It’s impossible to get rid of all faults Software and administrative errors still exist
Availability Strategy #2: Over-provisionStep 1: buy enough hardware to handle
your workloadStep 2: buy more hardware
Replicate
Replicate
Replicate
Replicate
Benefits of Replication
ScalabilityGuards against hardware failuresGuards against software failures (bugs)
Replication Meets Probability
p is probability that a single machine failsProbability of N failures is: 1-p^n
Siteunavailability
0.000001
0.00001
0.0001
0.001
0.01
0.1
1
0 1 2 3 4 5 6 7
Number of Replicas
Availability in the Real World
Phone network: 5 9’s 99.999% available
ATMs: 4 9’s 99.99% available
What about Internet services? Not very good…
2006: typical 97.48% Availability
97.48%97.48%
Source: Jim Gray
Netcraft’s Crisis-of-the-Day
What Gives?
Why isn’t simple redundancy enough to give very high availability?
Failure Modes
Fail-stop failure: A component fails by stopping It’s totally dead: doesn’t respond
to input or output Ideally, this happens fast
Like a light-bulb
Byzantine failure: Component fails in an arbitrary way Produces unpredictable output
Byzantine Generals
Basic goal: reach consensus in the presence of arbitrary failures
Results: More than 2/3 of the nodes must be “loyal”
3t + 1 nodes with t traitors Consensus is possible, but expensive
Lot’s of messages Many rounds of communication
In practice, people assume that failures are fail-stop, and hope for the best…
Example of a non Fail-Stop Failure
Server
Server
Server
Server
Server
Loadbalancer
Internet
Load Balancer uses a “Least Connections” policyServer fails by returning an HTTP error 400Net result: “failed” server becomes a black hole
Amazon.com
Correlated Failures
In practice, components often fail at the same time Natural disasters Security vulnerabilities Correlated manufacturing defects Human error…
Human errorHuman operator error is the leading cause of
dependability problems in many domains
Source: D. Patterson et al. Recovery Oriented Computing (ROC): Motivation, Definition, Techniques, and Case Studies, UC Berkeley Technical Report UCB//CSD-02-1175, March 2002.
59%22%
8%
11%
OperatorHardwareSoftwareOverload
51%
15%
34%
0%
Public Switched Telephone Network Average of 3 Internet Sites
Sources of Failure
Understanding Human Error
Administrator actions tend to involve many nodes at once: Upgrade from Apache 1.3 to Apache 2.0 Change the root DNS server Network / router misconfiguration
This can lead to (highly) correlated failures
Learning to Live with Failures
If we can’t prevent failures outright, how can we make their impact less severe?
Understanding availability: MTTF: Mean-time-to-failure MTTR: Mean-time-to-repair Availability = MTTR / (MTTR + MTTF)
Approximately MTTR / MTTF
Note: recovery timeis just as importantas failure time!
Summary
Large distributed systems are built from many flaky components Key challenge: don’t let component failures become
system failures Basic approach: throw lots of hardware at the
problem; hope everything doesn’t fail at once Try to decouple failures Try to avoid single points-of-failure Try to fail fast
Availability is affected as much by recovery time as by error frequency