Event Sourcing #PHPTour 2015 @MattKetmo

CQRS Event Sourcing Reactive DDD }Patterns to build & structure applications Not new buzzwords, exist for many years How can it help to scale horizontally? Today’s focus

Matthieu Moquet @MattKetmo @ BlaBlaCar for ~4 years

{Disclaimer Oriented for long lived apps.
 Overkill for RAD / CRUD / TODO apps. Use those patterns in moderation.
 There is no silver bullet. This talk aims to open your mind.
 Think different!

I/O

90% ViewModels / Flexible Eventually consistent 10% Validation / Business rules Coherence (ACID) Read Write Fast & Scale

Use case bank accounts Account #1 ! 1337€ Account #2 ! 42€ Account #3 ! 10€ + Transfer money

Account ! id user_id amount User ! id name ...

Facade (Controller) Models (ORM) DB

//  Open  new  account   $account  =  new  Account();   $account-­‐>setAmount(0,  'EUR');   $em-­‐>persist($account);   $em-­‐>flush();   ! //  Credit  100  EUR   $account-­‐>setAmount(100,  'EUR');   $em-­‐>flush();   ! //  Debit  20  EUR   $account-­‐>setAmount(80,  'EUR');   $em-­‐>flush();

is not the solution When user process an operation on his account his doesn’t directly update his total amount of money We need a service that translate our business needs CRUD

Facade (Controller) Models (ORM) DB Service Layer (Business)

interface  BankAccountService     {          /**  @return  int  */          public  function  open();   !        /**  @return  Account  */          public  function  get($accountId);   !        /**  @return  Account[]  */          public  function  findAll($userId);   !        /**  @return  void  */          public  function  credit($accountId,  $balance);   !        /**  @return  void  */          public  function  debit($accountId,  $balance);   }  

CQS Command Query Separation }Each operation is either a Command or a Query, not both

Commands Queries Change the state of the system. Do not return anything. Does not change the state of the system. Return data. Write only Read only

Commands /**  @return  Account  */   public  function  get($accountId);   ! /**  @return  Account[]  */   public  function  findAll($userId); /**  @return  int  */   public  function  open();   ! /**  @return  void  */   public  function  credit($accountId,  $balance);   ! /**  @return  void  */   public  function  debit($accountId,  $balance); Should be void Queries

/**      *  @param  Uuid  $accountId    *    *  @return  void      */   public  function  open(Uuid  $accountId);   Let the client generate the identifier, not the infrastructure (database)

CQRS Command Query Responsibility Segregation }Like CQS but separate queries and commands in different classes

Command Facade Query Facade Command Service Query Service Models (ORM) DB

CQRS challenges the assumption that reading and writing are sharing the same abstractions Databases Models Apps

Command Facade Query Facade Command Handler Query Repository Command Models DB Read Models Command DTO Segregation of read and write is a radical form of decoupling

No content

The interesting part about CQRS is not the pattern itself. But it allows to challenge established assumptions and opens new architectural options.

How to Scale

Now that the workflow of Read & Write has been separated. We can optimize each side separately.

asynchronously Since Command Handlers do NOT return anything, 
 then do the execution

Command Facade Command Handlers Commands Queueing Ack

! $accoundId  =  AccountId::generate();   $command  =  new  OpenAccountCommand($accountId);   ! if  (/*  is  command  valid  */)  {          $commandBus-­‐>handle($command);   !        return  new  Response(                  null,                  $async  ?  202  :  201,  //  Accepted  :  Created                  ['Location'  =>  "/accounts/$accountId"]          );     }

$command  =  new  DebitCommand(          $accountId,            new  Money(20,  'EUR')   );   ! if  (/*  account  has  enough  money  */)  {          $commandBus-­‐>handle($command);   !        return  new  Response(null,  204);   }

Command tracking Scale workers Horizontally by instantiating more workers (subscribers) Generate an identifier per command to let the client tracks its status Rabbitmq, Kafka, Gearman, ……… make your choice on queues.io

Projections On the Read side you should prepare you data as it will be requested

Having only 1 data store for reads & writes does not scale (well) different usage different needs

Denormalize your data UI View models Statistic models Search index models API Read models ... Read models are faster than JOIN

• search & aggregations capabilities • horizontal scaling architecture (sharding, replication) • denormalization • fast Secondary data store to read your data

{      "user_id":  133742      "from":  "Paris",      "to":  "Luxembourg",      "by":  [          "Reims",            "..."      ],      "date":  "2015-­‐05-­‐11",   } Input (mysql)

[{      "trip":  {          "from":  "Paris",          "to":  "Luxembourg",          "date":  "2015-­‐05-­‐11",          "..."      },      "user":  {          "name":  "John  D",          "age":  29,          "grade":  "beginner",          "..."      }   }] Output (Elasticsearch) denormalized searchable

No content

SELECT  tweets.*,  users.*      FROM  tweets   ! JOIN  users      ON  users.id  =  tweet.sender_id   ! JOIN  follows      ON  follows.followee_id  =  user.id   ! WHERE  follows.follower_id  =  $userId   ! ORDER  BY  tweets.time  DESC   LIMIT  100

No content

Tweets stream DB Aggs Timelines {      "user_id":  1234567890      "status":  "Hello  World"      "timestamp":  1430491773   } ! [{      "tweet_id":  1234567890876543,      "username":  "MattKetmo",      "name":  "Matthieu  Moquet",      "timestamp":  1430491773,      "status":  "Hello  World",   },  {      "tweet_id":  1234567890886445,      ...   }]  

SELECT  tweets      FROM  timelines    WHERE  timeline_id  =  $userId    LIMIT  500

PageViewEvents DB Aggs Increment counters /month Google Analytics /day /hour total {      "eventType":  "PageViewEvent"      "timestamp":  1430491773,      "ipAddress":  "12.34.56.78",      "sessionId":  "abcd1234567890",      "pageUrl":  "/hello-­‐world",      "..."   }

Using Cassandra you need to: ‣ know the read requests before creating your data models ‣ create as many tables (ie. KeySpaces) than you have views ‣ denormalize the data (no join allowed) ’s keypoints

C* C* C* Primary Data C* Read Data User post message Endpoint

Multi Data Center Replication + Sharding + Cache

Command Command Handler Primary DB Projections View Models Message Bus Write Read Eventual Consistency {...} {...} {...} {...}

Event Sourcing change the way we store our primary data

If we know the events of the past, we can reconstitute the present. Event Sourcing

Event Sourcing Register only a series of events. Reconstitute the state of current "entity" by reading the past events.

the standard way Something happen State A State B Something happen (delta) (dropped) (stored) (delta) Time

the event sourcing way Something happen State A State B Something happen (stored) (reconstituted) (reconstituted) (stored) Time

[{        "uuid":  "110e8400-­‐e29b-­‐11d4-­‐a716-­‐446655440000",        "type":  "AccountWasOpen",        "recorded_on":  "2015-­‐05-­‐11T13:37:00Z",        "payload":  {}   },  {        "uuid":  "110e8400-­‐e29b-­‐11d4-­‐a716-­‐446655440000",        "type":  "AccountWasCredited",        "recorded_on":  "2015-­‐05-­‐11T14:42:00Z",        "payload":  {  "amount":  100,  "currency":  "EUR"  }   },  {      "..."   }] Event Store

Aggregates Domain models which serve the business logic using the ubiquitous language. Work as an isolated object (or graph of objects) which doesn’t reference any others. Can ensure only its own integrity.

//  Open  a  new  account   $accountId  =  AccountId::generate();   $account  =  BankAccount::open($accountId);   ! //  Add  some  money   $account-­‐>credit(new  Money(100,  'EUR'));

Records events Aggregates class  BankAccount  implements  AggregateRootInterface   {          //  ...   }

interface  AggregateRoot   {          /**            *  @return  DomainEventStream            */          public  function  getUncommittedEvents();   !        /**            *  @return  string            */          public  function  getAggregateRootId();   }

$accountId  =  AccountId::generate();   $account  =  BankAccount::open($accountId);   $account-­‐>credit(new  Money(100,  'EUR'));   ! //  Retrieve  event  stream   $events  =  $account-­‐>getUncommittedEvents();   //  -­‐  AccountWasOpen   //  -­‐  AccountWasCredited

class  BankAccount  extends  EventSourcedAggregateRoot   {          private  $accountId;   !        public  function  getAggregateRootId()          {                  return  $this-­‐>accountId;          }   !        public  static  function  open(AccountId  $accountId)          {                  $account  =  new  self();                  $account-­‐>apply(new  AccountWasOpen($accountId));   !                return  $account;          }   !        public  function  credit(Money  $balance)          {                  $this-­‐>apply(new  AccountWasCredited(                          $this-­‐>accountId,                          $balance-­‐>getAmount(),                          $balance-­‐>getCurrency()                  );          }   }

Reconstitute from events Aggregates

class  BankAccount  extends  EventSourcedAggregateRoot   {          //  ...              protected  function  applyAccountWasOpen(AccountWasOpen  $event)          {                  $this-­‐>accountId  =  $event-­‐>getAccountId();          }   !        protected  function  applyAccountWasCredited(AccountWasCredited  $event)          {                  $this-­‐>amount  +=  $event-­‐>getBalance();          }   !        protected  function  applyAccountWasDebited(AccountWasDebited  $event)          {                  $this-­‐>amount  -­‐=  $event-­‐>getBalance();          }   }

Aggregates Protect invariants

class  BankAccount  extends  EventSourcedAggregateRoot   {          //  ...   !        public  function  debit(Money  $balance)          {                  if  ($this-­‐>amount  <  $balance-­‐>getAmount())  {                          throw  new  NotEnoughMoneyException(                                  'Cannot  debit  more  than  current  amount'                          );                  }   !                return  $this-­‐>apply(new  AccountWasCredited(                          $this-­‐>accountId,                          $balance-­‐>getAmount(),                          $balance-­‐>getCurrency()                  ));                        }   }

Command + Event Sourcing

Command User intent Events Event Store Apply changes on system Primary DB Transactional append

class  BankAccountCommandHandler  extends  CommandHandler   {          public  function  handleOpenAccount(OpenAccount  $command)          {                  $account  =  BankAccount::open($command-­‐>getAccountId());   !                $this-­‐>repository-­‐>save($account);          }   !        public  function  handleCreditAccount(CreditAccount  $command)          {                  $account  =  $this-­‐>repository-­‐>load($command-­‐>getAccountId());                  $account-­‐>credit($command-­‐>getBalance());   !                $this-­‐>repository-­‐>save($account);          }   }

class  EventSourcingRepository  implements  RepositoryInterface   {          public  function  load($id)          {                  try  {                          $events  =  $this-­‐>eventStore-­‐>load($id);   !                        return  $this-­‐>aggregateFactory-­‐>create($events);                  }  catch  (EventStreamNotFoundException  $e)  {                          throw  AggregateNotFoundException::create($id,  $e);                  }          }   !        public  function  save(AggregateRoot  $aggregate)          {                  $events  =  $aggregate-­‐>getUncommittedEvents();   !                $this-­‐>eventStore-­‐>append(                          $aggregate-­‐>getAggregateRootId(),                            $events                  );          }   }

Event Sourcing PROS

Append Only Very fast data update Deleting data = new events

Immutable Infinite caching

Sharding Aggregates are self-reliant and don’t have external references No Joins

Complete historical data Being able to replay history is a major benefit both technically and for the business BI team will love it

No Database migration when updating the data DBA will love it

Event Sourcing CONS Well, not really

Ensure consistency How to avoid concurrent updates?

Load aggregate from EventStore: 150€ Create AccountWasDebited event Append event in datastore Example Two debits of 100€ must not be accepted if only 150€ left Load aggregate from EventStore: 150€ Create AccountWasDebited event Append event in datastore Event must NOT be appended twice

abstract  class  EventSourcedAggregateRoot  implements  AggregateRootInterface   {          private  $uncommittedEvents  =  array();   !        private  $playhead  =  -­‐1;            /**          *  Applies  an  event.            *  The  event  is  added  to  the  list  of  uncommited  events.          */          public  function  apply($event)          {                  $this-­‐>playhead++;   !                $this-­‐>uncommittedEvents[]  =  DomainMessage::recordNow(                          $this-­‐>getAggregateRootId(),                          $this-­‐>playhead,                          new  Metadata(array()),                          $event                  );          }   }

DBAL EventStore (MySQL) CREATE  TABLE  EventStore  (   !    `id`                    INT(11)  AUTO_INCREMENT,      `uuid`                BINARY(16),      `playhead`        INT(11),      `type`                TEXT,      `payload`          TEXT,      `metadata`        TEXT,      `recorded_on`  DATETIME,   !    PRIMARY  KEY  (`id`),   !    UNIQUE  KEY  `UNIQ_PLAYHEAD`  (`uuid`,  `playhead`)   );

E_TOO_MANY_EVENTS When your Aggregates are "long lived" it may be slow to read the full history eg. BankAccount

Snapshotting is a solution

Snapshotting Store the state of the aggregate and read only the events created after

Snapshotting Snapshot Snapshot Events States

TL;DR

Command Handler Events Event Store Command Auditing Read Model Event Listener Event Bus Projector Write Read Business rules Historical data Read / search index Side effects User intent View models

Frameworks C# / Java nCQRS Fohjin NEventStore LiteCQRS Lokad.CQRS Agr.CQRS Axon Framework JDON PHP Broadway Predaddy EventCentric.Core litecqrs-php ...?

qandidate-labs/broadway qandidate-labs/broadway-demo

GetEventStore.com Event Sourcing oriented database

cqrs.nu #MustRead

dddinphp.org Google Groups

Thank You! Slides available at moquet.net/talks/phptour-2015 Leave feedbacks @MattKetmo