6.824 2015 Lecture 1: Introduction
Note: These lecture notes were slightly modified from the ones posted on the 6.824 course website from Spring 2015.
Distributed systems
What is a distributed system?
- multiple networked cooperating computers
- Example: Internet E-Mail, Athena file server, Google MapReduce, Dropbox, etc.
Why distribute?
- to connect physically separate entities
- to achieve security via physical isolation
- to tolerate faults via replication at separate sites
- to increase performance via parallel CPUs/mem/disk/net
...but:
- complex, hard to debug
- new classes of problems, e.g. partial failure (did he accept my e-mail?)
- Leslie Lamport: "A distributed system is one in which the failure of a
computer you didn't even know existed can render your own computer
unusable."
- Advice: don't distribute if a central system will work
Why take this course?
- interesting -- hard problems, non-obvious solutions
- active research area -- lots of progress + big unsolved problems
- used by real systems -- unlike 10 years ago -- driven by the rise of big Web sites
- hands-on -- you'll build a real system in the labs
Course structure
See the course website.
Course components
- Lectures about big ideas, papers, labs
- Readings: research papers as case studies
- please read papers before class
- paper for today: MapReduce paper
- each paper has a question for you to answer and one for you to ask (see web site)
- submit question & answer before class, one or two paragraphs
- Mid-term quiz in class, and final exam
- Labs: build increasingly sophisticated fault-tolerant services
- First lab is due on Monday
- Project: design and build a distributed system of your choice or the system we pose
in the last month of the course
- teams of two or three
- project meetings with course staff
- demo in last class meeting
Main topics
Example:
- a shared file system, so users can cooperate, like Dropbox
- but this lecture isn't about dropbox specifically
- just an example goal to get feel for distributed system problems
- lots of client computers
Architecture
- Choice of interfaces
- Monolithic file server?
- Block server(s) -> FS logic in clients?
- Separate naming + file servers?
- Separate FS + block servers?
- Single machine room or unified wide area system?
- Wide-area dramatically more difficult.
- Client/server or peer-to-peer?
- Interact w/ performance, security, fault behavior.
Implementation
- How do clients/servers communicate?
- Direct network communication is pretty painful
- Want to hide network stuff from application logic
- Most systems organize distribution with some structuring framework(s)
- RPC, RMI, DSM, MapReduce, etc.
Performance
- Distribution can hurt: network b/w and latency bottlenecks
- Lots of tricks, e.g. caching, threaded servers
- Distribution can help: parallelism, pick server near client
- Idea: scalable design
- We would like performance to scale linearly with the addition of machines
N x
servers -> N x
total performance
- Need a way to divide the load by N
- divide the state by N
- split by user
- split by file name
- "sharding" or "partitioning"
- Rarely perfect
->
only scales so far
- Global operations, e.g. search
- Load imbalance
- One very active user
- One very popular file
->
one server 100%, added servers mostly idle
-> N x
servers ->
1 x
performance
Fault tolerance
- Dropbox: ~10,000 servers; some fail
- Can I use my files if there's a failure?
- Some part of network, some set of servers
- Maybe: replicate the data on multiple servers
- Perhaps client sends every operation to both
- Maybe only needs to wait for one reply
- Opportunity: operate from two "replicas" independently if partitioned?
- Opportunity: can 2 servers yield 2x availability AND 2x performance?
Consistency
- Contract w/ apps/users about meaning of operations
- e.g. "read yields most recently written value"
- hard due to partial failure, replication/caching, concurrency
- Problem: keep replicas identical
- If one is down, it will miss operations
- Must be brought up to date after reboot
- If net is broken, both replicas maybe live, and see different ops
- Delete file, still visible via other replica
- "split brain" -- usually bad
- Problem: clients may see updates in different orders
- Due to caching or replication
- I make
grades.txt
unreadable, then TA writes grades to it
- What if the operations run in different order on different replicas?
- Consistency often hurts performance (communication, blocking)
- Many systems cut corners -- "relaxed consistency"
- Shifts burden to applications
Labs
Focus: fault tolerance and consistency -- central to distributed systems.
- lab 1: MapReduce
- labs 2/3/4: storage servers
- progressively more sophisticated (tolerate more kinds of faults)
- progressively harder too!
- patterned after real systems, e.g. MongoDB
- Lab 4 has core of a real-world design for 1000s of servers
What you'll learn from the labs:
- easy to listen to lecture / read paper and think you understand
- building forces you to really understand
- "I hear and I forget, I see and I remember, I do and I understand" (Confucius?)
- you'll have to do some design yourself
- we supply skeleton, requirements, and tests
- but we leave you substantial scope to solve problems your own way
- you'll get experience debugging distributed systems
Test cases simulate failure scenarios:
- distributed systems are tricky to debug: concurrency and failures
- many client and servers operating in parallel
- test cases make servers fail at the "most" inopportune time
- think first before starting to code!
- otherwise your solution will be a mess
- and/or, it will take you a lot of time
- code review
- learn from others
- judge other solutions
We've tried to ensure that the hard problems have to do w/ distributed systems:
- not e.g. fighting against language, libraries, etc.
- thus Go (type-safe, garbage collected, slick RPC library)
- thus fairly simple services (MapReduce, key/value store)
Lab 1: MapReduce
- help you get up to speed on Go and distributed programming
- first exposure to some fault tolerance
- motivation for better fault tolerance in later labs
- motivating app for many papers
- popular distributed programming framework
- many descendants frameworks
Computational model
- aimed at document processing
- split doc
-> K1 k, list<V1> values
- run
Map(K1 key, list<V1> values)
on each split -> list<K2, V2> kvps
- run
Reduce(K2 key, list<V2> values)
on each partition -> list<V2>
- merge result
- write a map function and reduce function
- framework takes care of parallelism, distribution, and fault tolerance
- some computations are not targeted, such as:
- anything that updates a document
Example: wc
- word count
- In Go's implementation, we have:
func Map(value string) *list.List
- the input is a split of the file
wc
is called on
- a split is just a partion of the file, as decided
by MapReduce's splitter (can be customized, etc.)
- returns a list of key-value pairs
- the key is the word (like 'pen')
- the value is 1 (to indicate 'pen' occurred once)
- Note: there will be multiple
<'pen', 1>
entries in the list
if 'pen' shows up more times
func Reduce(key string, values *list.List) string
- the input is a key and a list of (all? ) the values mapped to that key in the
Map()
phase
- so here, we would expect a
Reduce('pen', [1,1,1,1])
call if pen appeared 4 times in the
input file
- TODO: not clear if it's also possible to get three reduce calls as follows:
Reduce('pen', [1,1]) -> 2
+ Reduce('pen', [1,1]) -> 2
Reduce('pen', [2,2])
- the paper seems to indicate
Reduce
's return value is just a list of values
and so it seems that the association of those values with the key 'pen' in this
case would be lost, which would prevent the 3rd Reduce('pen')
call
Example: grep
- map phase
- master splits input in
M
partitions
- calls Map on each partition
map(partition) -> list(k1,v1)
- search partition for word
- produce a list with one item if word shows up,
nil
if not
- partition results among
R
reducers
- reduce phase
- Reduce job collects 1/R output from each Map job
- all map jobs have completed!
reduce(k1, v1) -> v2
- identity function:
v1
in, v1
out
- merge phase
Performance
- number of jobs:
M x R
map jobs
- how much speed up do we get on
N
machines?
- ideally:
N
- bottlenecks:
- stragglers
- network calls to collect a Reduce partition
- network calls to interact with FS
- disk I/O calls
Fault tolerance model
- master is not fault tolerant
- assumption: this single machine won't fail during running a MapReduce app
- but many workers, so have to handle their failures
- assumption: workers are fail stop
- they fail and stop (e.g., don't send garbled weird packets after a failure)
- they may reboot
What kinds of faults might we want to tolerate?
- network:
- lost packets
- duplicated packets
- temporary network failure
- server disconnected
- network partitioned
- server:
- server crash+restart (master versus worker?)
- server fails permanently (master versus worker?)
- all servers fail simultaneously -- power/earthquake
- bad case: crash mid-way through complex operation
- what happens if we fail in the middle of map or reduce?
- bugs -- but not in this course
- what happens when bug in map or reduce?
- same bug in Map over and over?
- management software kills app
- malice -- but not in this course
Tools for dealing with faults?
- retry -- e.g. if packet is lost, or server crash+restart
- packets (TCP) and MapReduce jobs
- may execute MapReduce job twice: must account for this
- replicate -- e.g. if one server or part of net has failed
- replace -- for long-term health
Retry jobs
- network falure: oops execute job twice
- ok for MapReduce, because
map()/reduce()
produces same output
map()/reduce()
are "functional" or "deterministic"
- how about intermediate files?
- atomic rename
- worker failure: may have executed job or not
- so, we may execute job more than once!
- but ok for MapReduce as long as
map()
and reduce()
functions are deterministic
- what would make
map() or reduce()
not deterministic?
- is executing a request twice in general ok?
- no. in fact, often not.
- unhappy customer if you execute one credit card transaction several times
- adding servers
- easy in MapReduce -- just tell master
- hard in general
- server may have lost state (need to get new state)
- server may have rebooted quickly
- may need to recognize that to bring server up to date
- server may have a new role after reboot (e.g., not the primary)
- these harder issues you would have to deal with to make the MapReduce master fault tolerant
- topic of later labs
Lab 1 code
The lab 1 app (see main/wc.go
):
- stubs for
map() and reduce()
- you fill them out to implement word count (wc)
- how would you write grep?
The lab 1 sequential implementation (see mapreduce/mapreduce.go
):
- demo:
run wc.go
- code walk through start with
RunSingle()
The lab 1 worker (see mapreduce/worker.go
):
- the remote procedure calls (RPCs) arguments and replies (see
mapreduce/common.go
).
- Server side of RPC
- RPC handlers have a particular signature
RunWorker
rpcs.Register
: register named handlers -- so Call() can find them
Listen
: create socket on which to listen for RPC requests
- for distributed implementation, replace "unix" w. "tcp"
- replace "me" with a
<dns,port>
tuple name
ServeConn
: runs in a separate thread (why?)
- serve RPC concurrently
- a RPC may block
- Client side of RPC
call()
(see common.go
)
- make an RPC
- lab code dials for each request
- typical code uses a network connection for several requests
- but, real must be prepared to redial anyway
- a network connection failure, doesn't imply a server failure!
- we also do this to introduce failure scenarios easily
- intermittent network failures
- just loosing the reply, but not the request
The lab 1 master (see mapreduce/master.go)
- You write it
- You will have to deal with distributing jobs
- You will have to deal with worker failures