Clustered Architecture Patterns for Scalability and Availability

Clustered Architecture Patterns:
Delivering Scalability and Availability
Orion Letizi – Terracotta
Co-Founder and Software
Engineer

Confidential – for information of designated recipient only. Copyright Terracotta 2006

Agenda

  Patterns from Years of Tier-Based Computing

  Network Attached Memory / JVM-level clustering

  Clustered Architecture Patterns

  In-depth Review Of The New “Examinator” Reference
Implementation


The Innocent Enterprise Application

  Within every innocent enterprise application lies a sleeping
monster

  There comes a time when every successful enterprise application
outgrows its single machine architecture

  Whether for high-availability or scalability or both, the adult
enterprise application must live on more than one application
server

  That's when the latent beast strikes...


The State Monster


The State Monster

  User-scoped session state
–  Conversational state across
requests

  Application state
–  Global data, caches, indexes, etc.


The State Monster

  At Walmart.com we started like everyone else: stateless + load-balanced
+ Oracle (24 cpus, 24GB)

  Grew up through distributed caching + partitioning + write behind

  We realized that “ilities” conflict
–  Scalability: avoid bottlenecks
–  Availability: write to disk ( and I/O bottleneck )
–  Simplicity: No copy-on-read / copy-on-write semantics (relentless tuning, bug fixing)

  And yet we needed a stateless runtime for safe operation
–  Start / stop any node regardless of workload
–  Cluster-wide reboot needed to be quick; could not wait for caches to warm

  The “ilities” clearly get affected by architecture direction and the stateless
model leads us down a precarious path


The Precarious Path: Our tools lead us astray

  Stateless load-balanced architecture ⇒ bottleneck on DB

  In-memory session replication ⇒ bottleneck on CPU, Memory

  Clustered DB cache ⇒ bottleneck on Memory, DB

  Memcache ⇒ bottleneck on server

  JMS-based replicated cache ⇒ bottleneck on network

  …Pushing the problem between our app tier CPU and the data
tier I/O


CRUD Piles Up…

  Types of clustering:
–  Load-balanced (non-partitioned) Scale Out
–  Partitioned Scale Out

  Both Trade-off Scalability or availability (usually by hand) in
different ways

scalability

availability

  …and everything we do forces the trade-offs


The Stateless Convention: Use The Database


The Stateless Convention: Push State to the Database

  DB becomes a bottleneck

  Not optimized for object data


The Stateless Convention: Push State to Peers


The Stateless Convention: Push State to Peers

  Push state everywhere gives HA, but doesn't scale

  Push conversational state to buddy can't survive buddy pair death or
full cluster restart. Doesn't help for application state.


The Stateless Convention: Push State to Client

  Highly scalable for conversation state UNLESS state is large

  Doesn't help with application state


The Stateless Convention: Eats Developer Brains


The Stateless Convention: Eats Developer Brains

  Pushing state out of app server
doesn't get rid of state. There
is no such thing as quot;statelessquot;

  Serialization/externalization:
–  Full-object graph serialization
–  breaks object identity
–  corrupts data and programming
model

  quot;with a little ingenuity...quot;--
seems fun/easy to try, but ends
up taking over the development
effort


Changing the Assumptions: JVM-level Clustering


What Terracotta Is

Definition

  From “The Definitive Guide…”
–  Terracotta is Java infrastructure software that allows you to scale your
application for use on as many computers as needed, without expensive
custom code or databases

  NAS but for memory (almost)
–  A server process for central storage and locking
–  A transparent client I/O driver
–  A network protocol

  Network Attached Memory (NAM)


What Terracotta Is

Terracotta Deployment Model

Your JVMs
1.  Uses Hotspot JVM
2.  TC libraries are jars and bootclassloader
3.  Terracotta Server is pure Java
TC Libraries 4.  Terracotta Stores all data to disk
5.  Clustering in the box
TCP/IP

Terracotta Terracotta
Server Server
JVM JVM

replication
disk disk


What Terracotta Is

Changes In 1 JVM Visible in Others

JVM JVM JVM

Terracotta


What Terracotta Is

…Like This

JVM JVM JVM

mutate ACK ACK

Establish
single update
Persistent
TCP broadcast
Connection

Terracotta


What Terracotta Is

Oh Yeah: Did I Mention It is JVM-Level?

Before Terracotta After Terracotta

ID Monitor
ID Monitor

Object Object

Primitive Method() Primitive Method()
Field (value) Field (value)

Object Object
Field (ref) Field (ref)

ID Monitor ID Monitor


Performance + Reliability + Simplicity

  10X throughput over conventional APIs
–  All Reads from memory (implicit locality)
–  All Writes are deltas-only
–  Statistics and heuristics (greedy locks)

  Terracotta Active Server Array
–  Simple form available today for enterprise customers
–  V1.0 GA this year

  Looks like Java to me (code like your mom used to make)
–  Normal access patterns: no check-out before view and check-in on commit
–  Code and test at a unit level without infrastructure intruding on app logic
–  Threads on multiple JVMs look like threads on the same JVM


What Terracotta Is

Batching and Ordering In Accordance With JMM

  Durable

  Transactional

  Clustered

  Shared Memory

  Use synchronized or util.concurrent or Terracotta Integration
Module to lock


What Terracotta Is

What TC Does for You

  Pure Java
–  Shares your objects and framework internals
–  Shares thread coordination

  Efficient Scale-out through heuristics and statistics (what is
resident where)

  Availability in the box


What Terracotta Is

Reduces Clustering to Tuning

  Always coherent

  Always on disk at in-memory speed (streaming, not seeking)

  Always Pure Java (java.*, synchronized, and util.concurrent)

  Changes needed usually for tuning


When to use Terracotta: Will It Work With My App?
(The Terracotta API, so to
speak)


When to Use Terracotta

Users and Use Cases

  Web Apps   Customer Service
–  Portals –  CRM apps
–  Social networking –  Rewards / affinity programs
–  Gaming, betting

  Financials   Telco / IT
–  Trading apps –  VOIP
–  Trading bus –  Server mgmt



Terracotta is a Platform (not an app server)

  How do you share objects
–  HashMap, CHM, EHCache, HashTable, TreeMap

  How do you coordinate access
–  Wait / Notify, synchronized, util.concurrent

  Where do you store the data
–  In Terracotta as objects, write-behind to DB, write-thru to DB, etc.

  How do you process async workloads
–  Queue of changes, eventing pattern, etc.



API: Using Patterns That Are Clusterable

  Conversation clustering / persistence

  Window onto large dataset

  Data as a service

  Cluster membership and management

  Change Notification

  Shared Metadata



Pattern: Conversation Clustering

  Set a conversation key (cookie)

  Value should be object graph
–  not primitives like string or byte array

  Load Balancer required
–  Sticky load balancing (layer 7 preferred,
or Apache mod_jk)
–  Ensures locality of reference

  No DB required

Ref. impl. code available now:
http://svn.terracotta.org/svn/forge/projects/exam/

Account Creation

  SignupController
–  UserValidator
–  RegistrationService (DefaultRegistrationService)

@RequestMapping(method = RequestMethod.POST)
public ModelAndView processSubmit(final HttpServletRequest request, @ModelAttribute final User user,
final BindingResult result, final SessionStatus status)
throws RegistrationException, ProtocolException {
validator.validate(user, result);
if (result.hasErrors()) {
return new ModelAndView(quot;registration/signupquot;);
} else {
// make sure that no clear text passwords are stored in the back-end
user.setAndEncodePassword(user.getPassword());

// hold the user's registration until it has been confirmed over email
this.service.holdForEmailConfirmation(user, UrlHelper.createAbsoluteUrl(request,
ConfirmController.class));
status.setComplete();
return new ModelAndView(quot;registration/successquot;, quot;userquot;, user);
}
}


Account Creation: RegistrationService

DefaultRegistrationService
public class DefaultRegistrationService implements RegistrationService {

private final Configuration freemarkerConfiguration;
private final UserService userService;
private final MailSender mailSender;
@Root
private final Map<String, Long> heldUsers = new ConcurrentHashMap<String, Long>();
//...

public void holdForEmailConfirmation(final User user, final String confirmationUrl)
throws RegistrationException {
// ... Validate args

final String uuidString = SecurityHelper.generateUniqueCode();
heldUsers.put(uuidString, user.getId());

// ... Code to send confirmation email
}
}


Account Creation

  Durable Java data structure as temporary storage for confirmation
codes

  Benefits
–  No round trip to the database
–  No need for caching
–  No cache freshness issues (no opportunity for arbitrage)
–  No impedance mismatch at O/R layer

  Issues
–  Must clean up data structure in a cluster-aware way (similar to HTTP Session
cleanup)

  Password reset function works the same way


AuthC/AuthZ

  Spring Security TIM
–  Authentication (AbstractAuthenticationToken) and authorization
(GrantedAuthorityImpl) tokens held in HTTP Session via Spring Security
<filter>
<filter-name>springSecurityFilterChain</filter-name>
<filter-class>org.springframework.web.filter.DelegatingFilterProxy</filter-class>
</filter>
<filter-mapping>
<filter-name>springSecurityFilterChain</filter-name>
<url-pattern>/*</url-pattern>
</filter-mapping>

–  Clustered via Terracotta session clustering module
–  Spring Security TIM handles all instrumentation configuration for objects in
HTTP session

  LoginController


AuthC/AuthZ: Spring Security


AuthC/AuthZ: LoginController
@Controller
@RequestMapping(quot;/login.doquot;)
public class LoginController {
@RequestMapping(method = RequestMethod.GET)
public ModelAndView loginHandler(final HttpServletRequest request,
final HttpServletResponse response, final HttpSession session) {
final ModelAndView result = new ModelAndView(quot;user/loginquot;);

final boolean loginError = request.getParameter(quot;login_errorquot;) != null;
if (loginError) {
result.addObject(quot;loginErrorquot;, loginError);

final String lastUser
= (String) session.getAttribute(AuthenticationProcessingFilter.SPRING_SECURITY_LAST_USERNAME_KEY);
final Throwable lastException
= (Throwable)session.getAttribute(AbstractProcessingFilter.SPRING_SECURITY_LAST_EXCEPTION_KEY);
if (lastUser != null) {
result.addObject(quot;lastUserquot;, lastUser);
}
if (lastException != null) {
result.addObject(quot;errorMessagequot;, lastException.getMessage());
}
}

return result;
}
}


Conversation State: Take The Exam

  Spring WebFlow
–  Workflow through examination process

  ExamService (CachingWrapperExamService)
–  Caches exam meta-data (questions, answers, sections, etc.)

  ExamSessionService
–  Manages the state of in-progress examinations using clustered:
–  ReentrantReadWriteLock
–  Map


Conversational State: List Available Exams

<view-state id=quot;chooseExamquot; view=quot;exam/listquot;>
<on-render>
<evaluate expression=quot;examService.getAllExams()quot; result=quot;viewScope.examsquot; />
</on-render>
<transition on=quot;selectExamquot; to=quot;examSelectedquot;>
<set name=quot;flowScope.examIdquot; value=quot;requestParameters.examIdquot;></set>
</transition>
<transition on=quot;returnquot; to=quot;returnquot; />
</view-state>


this.cacheManager = CacheManager.getInstance();
this.cache =
this.cacheManager.getEhcache(EXAM_CACHE_NAME);
}

Conversational State: View Exam Details
<view-state id=quot;examSelectedquot; view=quot;exam/detailsquot;>
<on-render>
<evaluate expression=quot;examService.findById(flowScope.examId)quot; result=quot;viewScope.examquot;/>
</on-render>
<transition on=quot;backquot; to=quot;chooseExamquot; />
<transition on=quot;returnquot; to=quot;returnquot; />
<transition on=quot;startExamquot; to=quot;startExamquot; />
</view-state>

public class CachingWrapperExamService implements ExamService {
/* as defined in ehcache.xml */
private static final String EXAM_CACHE_NAME = quot;examCachequot;;

private final CacheManager cacheManager;
private final Ehcache cache;
private final ExamService examService;

public CachingWrapperExamService(ExamService examService){
this.examService = examService;
this.cacheManager = CacheManager.getInstance();
this.cache = this.cacheManager.getEhcache(EXAM_CACHE_NAME);
}
// ...
public Exam findById(Long id) {
Exam exam = getCached(id);
if (exam == null){
exam = cache(this.examService.findById(id));
}
return exam;
}
// ...
}


Conversation State: Start Exam
<action-state id=quot;startExamquot;>
<on-entry>
<evaluate expression=quot;examSessionService.startExam(currentUser.name, requestParameters.examId)quot;/>
</on-entry>
<evaluate expression=quot;examSessionService.getFirstQuestion(currentUser.name).getId()quot;
result=quot;flowScope.questionIdquot;/>
<transition to=quot;showQuestionquot;/>
</action-state>

public class ExamSessionServiceImpl implements ExamSessionService {

private static final Logger logger = Logger.getLogger(ExamSessionServiceImpl.class);
// a map storing userName -> examSession mapping of currently active exams
@Root
private final Map<String, ExamSession> ongoingExams = new HashMap<String, ExamSession>();
private final ScheduledExecutorService examTimeoutExecutor = Executors.newScheduledThreadPool(1);
private final Map<String, Future> scheduledTimeOutTasks = new HashMap<String, Future>();
@Root
private final ReentrantReadWriteLock rwLock = new ReentrantReadWriteLock();

@Autowired
public ExamSessionServiceImpl(final ExamService examService, final UserService userService) {
this.examService = examService;
this.userService = userService;
}

// ...
}


Conversation State: Start Exam (cont.)

// ...
public ExamSession startExam(final String userName, final Long examId) throws ExamException {
final User user = userService.findByUserName(userName);
ExamSession prev = null;
try {
rwLock.writeLock().lock();
prev = ongoingExams.get(userName);
if (prev != null &&
prev.getExam().getId().longValue() != examId.longValue()) {
throw new ExamAlreadyInProgressException(userName, prev.getExam(), quot;Exam already in progress for user: quot;
+ userName);
}
// return previously ongoing exam if attempted to start same exam again
if (prev != null) return prev;

final Exam exam = examService.findById(examId);
final ExamSession session = new ExamSession();
session.setUser(user);
session.setExam(exam);
ongoingExams.put(userName, session);

// run the timeout after 2 secs from the actual time so that exam time out happens on the client first
// and give the user a chance to get ExamTimedOutException when exam times out on client
final Future timeoutTask = examTimeoutExecutor.schedule(new ExamTimeoutTask(this, userName), session
.getRemainingTimeInSeconds() + 2, TimeUnit.SECONDS);
scheduledTimeOutTasks.put(userName, timeoutTask);
return session;
} finally {
rwLock.writeLock().unlock();
}
}
// ...
}

Coherent View Of Global Data: View Ongoing Exams
@Controller
@RequestMapping(quot;/exam/ongoing.doquot;)
@RolesAllowed( { StandardAuthoritiesService.ADMINISTRATOR })
public class OngoingExamsController {

private final ExamSessionService service;

@Autowired
public OngoingExamsController(final ExamSessionService examSessionService) {
this.service = examSessionService;
}

@RequestMapping(method = RequestMethod.GET)
public ModelAndView listOngoingExams() {
final ModelAndView result = new ModelAndView(quot;exam/ongoingquot;);
result.addObject(quot;examSessionsquot;, service.getOngoingExams());
return result;
}

protected OngoingExamsController() {
// protected default constructor is needed for CGLib AOPx
service = null;
}
}


Coherent View Of Global Data: View Ongoing Exams

private static final Logger logger = Logger.getLogger(ExamSessionServiceImpl.class);
// a map storing userName -> examSession mapping of currently active exams
@Root
private final Map<String, ExamSession> ongoingExams = new HashMap<String, ExamSession>();
private final ScheduledExecutorService examTimeoutExecutor = Executors.newScheduledThreadPool(1);
private final Map<String, Future> scheduledTimeOutTasks = new HashMap<String, Future>();
@Root
private final ReentrantReadWriteLock rwLock = new ReentrantReadWriteLock();

// ...

public Collection<ExamSession> getOngoingExams() {
Collection<ExamSession> values = null;
try {
rwLock.readLock().lock();
values = new ArrayList(ongoingExams.values());
} finally {
rwLock.readLock().unlock();
}
return values;
}

// ...

}


Conversation State And Coherent View Of Global Data

  Conversation State
–  Memory-speed access to flow data
–  Coherent cluster-wide
–  Database-like durability for failover
–  Handles failure of any/every node
–  Completely transparent: no put-back on change

  Coherent View Of Global Data
–  Coherent cluster-wide
–  No round-trip to database
»  Ergo, no database abuse
»  Ergo, no caching
»  Ergo, no cache freshness probs
–  Memory-speed read locks
–  Simple data model
»  ReentrantReadWriteLock for global thread access control
»  Map holds global data
–  Need to refine large data set access pattern
»  @ 20K ongoing exams, single-page access is non-sensical

Case Study: Conversation Clustering


Large Publisher Gets Caught Down the Path with Oracle

Scaling Out or Up?


Breaking the Pattern without Leaving “Load-Balanced” World
• $1.5 Million DB & HW savings
• Doubled business
• More than halved database load


Results

  Database was still the SoR which kept reporting and backup
simple

  Scalability was increased by over 10X

  Availability was not compromised since test data was still on disk,
but in memory-resident format instead of relational

  …simple scalability + availability



Pattern: Window Onto Large Dataset

  Isolate the JVMs from each other carefully
–  ConcurrentHashMap of ConcurrentHashMaps
–  One map per L1 JVM

  If you must lock while updating, lock the value graph itself (fine-
grained) using ReadWriteLock

  Enough Heap in each JVM to hold most data it needs
–  High pre-fetch rate

  You must come up with a way to load balance or partition or this
will thrash the network



Pattern: Data As A Service (impl coming December)

  Service Oriented

  Write a stateless get() router with ActiveMQ
–  Send all messages to the get() router
–  He hashes and sends the get() to appropriate stripe
–  Transparent to service consumer

  Fairly flat key / value pairs
–  String or Integer key
–  Value == shallow object graph (mostly primitives)

  Key must hash evenly. Don’t hash spatially even though it is
tempting. (e.g., “New York customers” is worse than “A”)

  Use MySQL to index keyspace and execute queries that return
vectors of key/value pairs to load

  Update MySQL write-behind (another pattern)



Pattern: Cluster Membership

  Want to know when L1 JVMs come and go
–  Listen for Terracotta JMX events
–  Catch L1 disconnect event and deterministically take over work (next worker
ID after failed worker)

  Want to know when L2 JVMs fail over
–  Listen for Terracotta JMX events
–  If you don’t reconnect after certain time, system.exit() and use initd to restart
yourself
–  Avoid disconnected mode

  Always use a surviving L1 JVM to address a departing one. Do
not use System exit hooks! Clean up on restart if last JVM in the
cluster

  Don’t requeue work, steal queue references instead



Pattern: Change Notification (impl coming soon)

  Use this one for write-behind

  Linked Blocking Queue in each L1 JVM

  Share the LBQ via DSO in a CHM of queues

  Have your changeable objects implement an iChangeable
interface. Call flush() method on change. You will have to
implement this method

  Daemon thread takes batches of object references and calls flush
on each.

  Idempotency is req’d to avoid transactions / JTA. Then peek
instead of take and take after flush().

  Use Cluster membership (another pattern) to recover from L1 JVM
failure


Pattern: Shared Metadata

  Composite Pattern…

  Use EHCache + Linked Blocking Queue

  Conversation Clustering pattern

  Change Notification pattern


Summary
  Simplicity, scalability, and availability can be friends

  Write normal Java code that works across JVMs

  Use clustered architecture patterns
–  Conversation clustering/persistence
–  Window onto large dataset
–  Change notification
–  Shared metadata
–  Clustered membership and data management

  Leave business data in the database, use durable NAM for
application state data

  Centralized operational control: manage your application cluster
like you do your database

  Terracotta is open source
–  Free to use through production
–  Enterprise subscription, training, and services available

Resources

  Open Source (MPL-based) JVM-level clustering:
http://www.terracotta.org

  Apress / Amazon.com: “Definitive Guide to Terracotta”
–  By Alex Miller, Ari Zilka, Geert Bevin, Jonas Bonér, Orion Letizi, Taylor Gautier

  Forums: http://forums.terracotta.org/
  Enterprise Offerings: http://www.terracottatech.com/


Clustered Architecture Patterns for Scalability and Availability

Recomendados

Recomendados

Mais conteúdo relacionado

Semelhante a Clustered Architecture Patterns for Scalability and Availability

Semelhante a Clustered Architecture Patterns for Scalability and Availability (20)

Último

Último (20)

Clustered Architecture Patterns for Scalability and Availability