Wednesday, January 16, 2013

Overengineering

Empty layers of abstraction that are at best unnecessary and at worst restrict you to a narrow, inefficient use of the underlying API. Michael Borgwardt
In the cases where we've considered things over engineered, it's always been describing software that has been designed to be so generic that it loses sight of the main task that it was initially designed to perform, and has therefore become not only hard to use, but fundamentally unintelligent. Darasd
It fascinates me how some people are willing to spend so much time and brainpower on the so-called "application infrastructure" they unknowingly replicate, rewrite existing frameworks--mature ones, even those they are already using; they turn a blind eye to economy of execution, development productivity, and the fact that these things often take longer than developing parts that actually fulfill the requirements.

I've recently seen many cases of over-engineering and over-tinkering.  For example my superior was recently fretting about object casting.  So basically we store our project configuration in a database table, and these configuration parameters are sometimes used as method arguments.  They also store the data type name for each parameter in the table.  And then they already do the casting during query result mapping.  And for some reason, they have a global Map of String and Object to retain the result, completely making all of that casting useless, and to use it as a method argument they have to cast it again from Object into the required type.  And he wanted to come up with generics and a layer of abstraction on top of, well, what it is I don't have the slightest clue--to "avoid object casting".  I was completely speechless.  My rationale is that object casting is unavoidable when you: 1. keep them all, of different types, as values of a single Map, 2. store them as strings in the database, and 3. use Java (in Scala you can do a type matching to avoid explicit object casting).  Although, well, now that I think about it again, this is probably not a case of over-engineering but probably leans more towards a case of uninformed engineering.

Friday, December 7, 2012

My Cassandra Templates

So I thought I would share some of my abstract adapter code for, eh, interacting with Hector client, that in turn interacts with Cassandra.

The idea is it adapts any model/entity POJO class into a key and columns that fit a column family that corresponds to the model. This happens through the marshal and unmarshal methods that are specific to each implementor of this trait. Curious names--marshal and unmarshal, but I just don't want to call them map and unmap or pack and unpack because in Scala they have different concepts. Marshal transforms the model into, really, a Scala tuple of key and list of tuples of column name and value. The persist methods in this adapter know how to work with them. Unmarshal transforms query results (in the form of ColumnFamilyResult) into the model. And this one is used by query methods.

As you can see the adapter also functions as a high-level DAO in that the user of the DAO shouldn't know about how to work with Hector/Cassandra.

Fine, I know it gets roundabout, the interaction between this trait and its subclass, but it's just how it is for now. I don't want to spend too much time refining for now.

Furthermore there is this automatic Id field finder method (getKeyFieldOpt). It does some reflection work against the model to find which field is the Id (row key). In my case Id is a case class (like var id = Id(9999)). And I think I use some implicit defs so retrieving id transforms it automatically to its value whenever relevant. But you get the gist.

This is a work in progress. At the moment all generic queries work to fetch the entire row of a column family and unmarshal its data to a model, but I'll be working on generic slices and ranges soon.

So here's the base trait:

import java.lang.reflect.Field
import scala.collection.JavaConversions._
import id.pronimo.watchlet.exception.NoExistingRowException
import id.pronimo.watchlet.exception.WrongModelException
import javax.validation.ConstraintViolation
import javax.validation.ConstraintViolationException
import javax.validation.Validation
import javax.validation.ValidationException
import me.prettyprint.cassandra.serializers.AsciiSerializer
import me.prettyprint.cassandra.serializers.StringSerializer
import me.prettyprint.cassandra.serializers.UUIDSerializer
import me.prettyprint.cassandra.serializers.CompositeSerializer
import me.prettyprint.cassandra.serializers.DateSerializer
import me.prettyprint.cassandra.service.template.ColumnFamilyResult
import me.prettyprint.cassandra.service.template.ColumnFamilyTemplate
import me.prettyprint.cassandra.service.template.ColumnFamilyUpdater
import me.prettyprint.hector.api.exceptions.HectorException
import me.prettyprint.hector.api.factory.HFactory
import me.prettyprint.hector.api.mutation.Mutator
import me.prettyprint.hector.api.Keyspace
import id.pronimo.watchlet.model.metadata.Id
import me.prettyprint.cassandra.serializers.SerializerTypeInferer
import me.prettyprint.hector.api.Serializer

trait CommonStandardCFAdapter[R <: Serializable, K, N] {
 @transient val (us, ss, as, ds, cs) = (UUIDSerializer.get, StringSerializer.get, AsciiSerializer.get, DateSerializer.get, CompositeSerializer.get)
 @transient val EMPTY_BYTE_ARRAY = Array.empty[Byte]
 @transient val validator = Validation.buildDefaultValidatorFactory.getValidator
 
 def fs[F](x:F) = { SerializerTypeInferer.getSerializer[F](x) }
 val hcolumn = (n:N, v:Any) => {
  HFactory.createColumn(n, v, getKeyspace.createClock, getNameSerializer, fs(v))
 }

 /* CONCRETE METHODS [BEGIN] */
 /**
  * Add a new record, or override existing data.
  *
  *  @tparam   R the record type defined per implementation
  */
 @throws(classOf[HectorException])
 @throws(classOf[ValidationException])
 def write(record: R, deleteFirst: Boolean) {
  validate(record)
  if (deleteFirst) remove(record)
  val thriftTemplate = this.getThriftTemplate
  val (k, c) = marshal(record)
  val updater = thriftTemplate.createUpdater(k)
  c.iterator.foreach {
   case (n, v) =>
    updater.setColumn(hcolumn(n, v))
  }
  thriftTemplate.update(updater)
 }

 /**
  * Update a record by specifying intended column to add/update.
  *
  * Will throw an exception when record with specified key does not exist.
  *
  *  @note  Use with caution, this method allows adding a column not defined in the record type.
  *  @param    key of the record
  *  @param    colName the column name
  *  @param    value the column value
  *  @tparam   K the key type defined per implementation
  *  @tparam   V the value type
  */
 @throws(classOf[HectorException])
 @throws(classOf[ValidationException])
 def update[V](key: K, colName: N, value: V) {
  val thriftTemplate = this.getThriftTemplate

  if (false == thriftTemplate.isColumnsExist(key)) {
   throw new NoExistingRowException(key, "update")
  }
  val updater = thriftTemplate.createUpdater(key)

  val column = HFactory.createColumn(colName, value, getKeyspace.createClock)
  updater.setColumn(column)
  thriftTemplate.update(updater)
 }

 /**
  * Update a record by specifying multiple intended columns to add/update.
  *
  * Will throw an exception when record with specified key does not exist.
  *
  *  @note  Use with caution, this method allows adding columns not defined in the record type/column family metadata.
  *  @param    key of the record
  *  @param    map of column name/column value
  *  @tparam   K the key type defined per implementation
  */
 @throws(classOf[HectorException])
 @throws(classOf[ValidationException])
 def update(key: K, map: Map[N, Any]) {
  val thriftTemplate = this.getThriftTemplate

  if (false == thriftTemplate.isColumnsExist(key)) {
   throw new NoExistingRowException(key, "update")
  }

  val updater = {
   thriftTemplate.createUpdater(key)
  }
  map.iterator.foreach {
   case (n, v) =>
    val column = HFactory.createColumn(n, v, getKeyspace.createClock)
    updater.setColumn(column)
  }
  thriftTemplate.update(updater)
 }

 /**
  * Add multiple records, or override when such with the same key exists.
  *
  * Will throw an exception when record with specified key does not exist, automatically cancelling whole operation.
  *
  *  @note  Use with caution, this method allows adding columns not defined in the record type/column family metadata.
  *  @param    key of the record
  *  @param    map of column name/column value
  *  @tparam   K the key type defined per implementation
  */
 @throws(classOf[HectorException])
 @throws(classOf[ValidationException])
 def writeBatch(records: List[R], deleteFirst: Boolean) {
  val cf = this.getThriftTemplate.getColumnFamily
  val thriftTemplate = this.getThriftTemplate

  records.iterator.foreach { it => validate(it) }

  val mutator = thriftTemplate.createMutator

  records.iterator.foreach { it =>
   if (deleteFirst) remove(it)
   marshal(it) match {
    case (k, c) => c.iterator.foreach { tuple =>
     mutator.addInsertion(k, cf, HFactory.createColumn(tuple._1, tuple._2, getKeyspace.createClock))
    }
   }
  }
  thriftTemplate.executeBatch(mutator)
 }

 /**
  * Read record by key, returning Option
  *
  *  @param    key of the record
  *  @tparam   K the key type defined per implementation
  *  @tparam   R the record type defined per implementation
  */
 @throws(classOf[HectorException])
 def read(key: K): Option[R] = {
  val thriftTemplate = this.getThriftTemplate
  val result = thriftTemplate.queryColumns(key)
  unmarshal(result)
 }

 /**
  * Remove entire row by key
  *
  *  @param    key of the record
  *  @tparam   K the key type defined per implementation
  */
 @throws(classOf[HectorException])
 def remove(key: K) {
  this.getThriftTemplate().deleteRow(key)
 }

 /**
  * Remove entire row using the record's id
  *
  *  @tparam   R the record type defined per impl
  *  ementation
  */
 @throws(classOf[HectorException])
 def remove(record: R) {
  remove(getKey(record))
 }

 /**
  * Get key field option (thru reflection)
  */
 protected def getKeyFieldOpt[E <: R: Manifest] = implicitly[Manifest[E]].erasure.getDeclaredFields.find(_.getType.equals(classOf[Id[K]]))
 
 /**
  * Get N serializer (thru reflection)
  */
 protected def getNSerializer[E <: N: Manifest]():Serializer[N] = SerializerTypeInferer.getSerializer[N](implicitly[Manifest[E]].erasure)

 /**
  * Get key for record (thru reflection)
  */
 def getKey[E <: R: Manifest](record: R): K = {
  try { getKeyField.get(record).asInstanceOf[Id[K]]}
  catch {
   case e =>
    throw new WrongModelException(implicitly[Manifest[E]].erasure, "Key type specified in template does not match actual key type of record.", e)
  }
 }

 /**
  * Validate record using JSR-303
  *
  *  @param    record
  *  @tparam   R the record type defined per implementation
  */
 @throws(classOf[ValidationException])
 def validate(record: R) {
  val violations = validator.validate(record)

  if (false == violations.isEmpty()) {
   val exceptionMessage = violations.iterator
    .map(it => it.getPropertyPath.toString + " " + it.getMessage)
    .mkString(", ")

   throw new ConstraintViolationException(exceptionMessage, violations.asInstanceOf[java.util.Set[ConstraintViolation[_]]])
  }
 }
 /* CONCRETE METHODS [END] */

 /* ABSTRACT METHODS [BEGIN] */
 /**
  * Transform record into tuples with format (key, List[(columnName, columnValue)])
  */
 def marshal(record: R): (K, List[(N, Any)])

 /**
  * Transform ColumnFamilyResult into record
  */
 def unmarshal(columns: ColumnFamilyResult[K, N]): Option[R]

 /**
  * Get underlying Thrift template used by this spec template
  */
 def getThriftTemplate(): ColumnFamilyTemplate[K, N]
 
 /**
  * Get column name serializer used by this spec template
  * This became mandatory for creating composite columns using the Hector client.
  */
 def getNameSerializer():Serializer[N]

 def getKeyField: Field

 def getKeyspace: Keyspace
 /* ABSTRACT METHODS [END] */
}

And a sample implementation:
PS: Don't mind the CDI annotations. And the transient annotations are there just so it plays along well with CDI.

@Named( "ItemRepository" )
@ApplicationScoped
class ItemCFAdapter extends CommonStandardCFAdapter[Item, UUID, String] with Serializable {
 @transient val keyField = {
  getKeyFieldOpt[Item] match {
   case Some( field ) =>
    field.setAccessible( true )
    field
   case None =>
    throw new WrongModelException( classOf[Item], "Cannot find field with Id metadata." )
  }
 }

 @transient val nameSerializer = getNSerializer[String]

 @Inject
 @transient
 var keyspace: Keyspace = _

 var thriftTemplate: ThriftColumnFamilyTemplate[UUID, String] = _

 @PostConstruct
 def initialise() {
  thriftTemplate = new ThriftColumnFamilyTemplate( keyspace, CF.CF_NAME, us, as )
 }

 override def marshal( record: Item ): ( UUID, List[( String, Any )] ) = {
  ( getKey( record ), List(
   ( "label", record.label ),
   ( "url", record.url ) ) )
 }

 override def unmarshal( columns: ColumnFamilyResult[UUID, String] ): Option[Item] = {
  val l = columns.getString( "label" )
  val u = columns.getString( "url" )
  ( Option( l ), Option( u ) ) match {
   case ( Some( _ ), Some( _ ) ) =>
    Option( new Item {
     id = columns.getKey
     label = l
     url = u    } )
   case _ =>
    None
  }
 }

 override def getThriftTemplate() = thriftTemplate

 override def getKeyField() = keyField

 override def getNameSerializer() = nameSerializer

 override def getKeyspace() = keyspace

}

Thursday, November 29, 2012

Back to JSP

Alright, so JSP is old and perhaps no longer endorsed by Oracle (there's no section on JSP in the Java EE 6 tutorial).  But it, I find, remains a sane, mature, and better-supported choice of technology for generating web content.  Many other templating engines suffer from lack of support and continued development (due to less usage), crippled integration with other technologies (i.e. frameworks), or unsatisfactory tooling.

Most of the hate-bordering rants on JSP that I see have nothing to do with the JSP technology itself.  The problems they face arise from anti-patterns: spaghetti code, lack of separation of concern, or simply outdated version of JSP spec.

What would be cool for the next JSP spec update is a way to supply partial XML directly from the controller, like that in Lift web framework.  With CDI, yeah.  Something like:

@Produces
@XmlNode
@Named("calendar")
def getCalendar:XmlNode = {
  return 

Calendar

Monday to Sunday
}

Saturday, November 10, 2012

JAX-RS ftw

I never realised this before, partly because the Java EE 6 documentation doesn't make it obvious, but apparently one can use the JAX-RS to create web page controllers.  It seems to be implementation specific, however.  JBoss Resteasy, with Seam Resteasy, has built-in support for freemarker and Velocity templates as well as undocumented for JSP.  Jersey, on the other hand, afaik, only does for JSP.  It's pretty snazzy considering the features JAX-RS has and its deep integration with the Java EE stack--all those REST methods, CDI scopes, streamlined authentication and authorization, etc.
Bottom line, Java EE 6 hasn't abandoned stateless web!  Or something like that.

Sunday, October 28, 2012

My Little Take on Backend Development

Update 29-10-2012:
I found this blog post more comprehensively in defense of Java EE 6.
Original post:
This comes from someone who only just recently read about JTA and JTS. Take this with a grain of salt, so to speak.

I have some experience using the Spring Framework, mainly the Spring MVC sometime about 2years ago, and some again just recently (as indicated in my previous post).  Back then, it was first a multifinancing application, second a message broadcasting engine, sort of.  Neither of which I created from the ground up.  I was merely adding features and so on.  I couldn't compare it against the Java EE 6 at that time, lack of understanding of their internals and so on.

The message broadcasting engine had to be improved so that it could support more load.  There were topics such as clustering, high availability, parallelism, concurrency, etc.  I think it was then that I came across the JMS and some asynchronous processing features provided by the EJB 3.  That was to be my gateway to Java EE 6.   Afterwards it was a haphazard series of choice to learn the actor model, CSP, Scala, Erlang, Haskell, NoSQL, etc.

So I wonder what was it again, Spring Framework's winning points.  Their documentation is among the best.  Their framework itself is rock solid, for what it is.  What it is, really, is a little confusing, though.  It appears to be sort of an alternative to the Java EE stack, but heavily relies on it.  It manages the object lifecycles.  Java EE does that as well, manage object lifecycles.  Two things manage your object lifecycles.  Obviously you can run Spring Framework on a standard servlet container, but I see everywhere that it is recommended that you use, say, the JTA provided by the container when available.  It would be unavoidable that both containers will manage the object lifecycles if you used CMT alongside Spring.  And how is it that one could share a bean across deployments in Spring?  Even if you could, I think using Spring remoting, your object lifecycles management would be more complicated because it would be a communication between two unrelated containers.  It looks to me that the EJB specification is superior in this regard, with its local and remote beans concept and tighter integration with JTA and the rest of the Java EE specification.

Anyway, honestly I'm not a big fan of dependency injection, Spring or CDI.  It's a little too much magic for me.  Language level singleton object construction available in Scala and programmatic lookup and vending are easier to work with.

Friday, October 26, 2012

Integrating Spring Framework, Jetty, and JBoss Narayana

Update 29-10-2012:
This does not work. I posted this only after testing rollback in a transaction, and it worked. However, I couldn't get it to work as Hibernate's transaction manager alongside Spring; e.g. update operations miss the transaction scope. I tried many setup variations. If someone manages to get this working, it'd be great if you shared. For the time being I'm falling back to CMT provided by JBoss AS.
Original post:
I had to browse through quite many websites and tested some adaptations to eventually get this to work. Shame on me, it's not all that complex how it ends up.
Oh, and my syntax highlighter does this XML tag uppercase thing that is so uncalled for.
My stack goes as following:
  • Spring Framework 3.1.2.RELEASE
  • JBoss Narayana 4.17.1.Final
  • Apache Derby (network) DB 10.9.1.0
  • Jetty 8.1.7.v20120910
The Spring Framework of this version doesn't even need persistence.xml to be present, so configuration, linking data source, persistence unit, and transaction manager, in a single Spring context XML is possible and makes more sense.  The JBoss Narayana JTA guide shows that to incorporate JDBC connections within itself:
  1. Use the com.arjuna.ats.jdbc.TransactionalDriver that comes with the Narayana library
  2. Wrap a javax.sql.XADataSource by creating an implementation class of com.arjuna.ats.internal.jdbc.DynamicClass and bind the data source to a JNDI (programmatically).
    The TransactionalDriver will use the DynamicClass as its data source provider.
I stumbled at binding.  Admittedly I'm not quite sure how it works on Jetty.  I managed to make it work by, firstly, bind the XADataSource to JNDI using Jetty's context configuration. Here's how it looks like in my jetty-env.xml.

 
 jdbc/dbxa
 
  
   mydb
   username
   password
  
 

Secondly, I let my implementation of DynamicClass, in its concrete getDataSource(dbName:String):XADataSource method to perform JNDI lookup, cast it to XADataSource, and return it. Simple as that. Why did I even bother struggling with Spring JNDI binding and JMX and whatever else. Here's how it looks like:
public XADataSource getDataSource(String dbName, boolean create)
  throws SQLException {
 try {
  InitialContext ic = new InitialContext();
  ClientXADataSource40 xa = (ClientXADataSource40) ic
    .lookup("jdbc/dbxa");
  return xa;
 } catch (NamingException e) {
  throw new SQLException(null, e);
 }
}
Afterwards, I bind things together in the Spring context XML.
Since I already defined the credentials in the Jetty context configuration, I don't need to repeat it here. The driver properties bean is just:

 
  
   my.implementation.of.DynamicClass
  
 

Next is the data source bean. I have to prepend the JNDI with jdbc:arjuna:, that's just how it works.

 
  com.arjuna.ats.jdbc.TransactionalDriver
 
 
 
  
 

And then the EntityManagerFactory bean, which also replaces persistence.xml. I happen to use Hibernate, so that's HibernateJpaVendorAdapter for me.

 
 
 
 
  
   
   
   
  
 
 
 
  
 

I also need to define the Narayana transaction manager and user transaction implementation beans as well as the Spring transaction manager bean itself.




 
  
 
 
  
 

And lastly if you like annotations:
As a bonus, you can also set where Narayana will store its objects, default timeout, etc by defining this factory bean.

 
 
 
  
   60
   .\arjuna
   .\arjuna
  
 

That's all. You can add a logger for com.arjuna.ats and set it to trace or fine to see whether it works. I don't suppose I need to show an example of how transactional annotation is used right. And there's that org.springframework.transaction.support.TransactionTemplate and org.springframework.transaction.interceptor.TransactionAspectSupport to allow manual, programmatic transaction control.

Saturday, October 20, 2012

Lift Web Framework is Easy

It's ubiquituously said that the Lift web framework has a steep learning curve.  Hard to believe when

<div class="lift:surround?with=default;at=content" id="real_content">
 <h1>Welcome to your project!</h1>
 <lift:Hello.world />  
</div>
object Hello extends DispatchSnippet {
 val dispatch: DispatchIt = {
  case name => render(name) _
 }
 def render(name: String)(ignore: NodeSeq): NodeSeq = Text("Hello, " + name)
}
LiftRules.snippetDispatch.append {
 case "Hello" => id.openfx.openshop.snippet.Hello
}
simply wraps the page with the template 'default.html' and makes a snippet Hello.world lookup.

Everything in Lift is easy and transparent, I daresay. Most things can be reasoned as you write code, instead of having to read its specification. I can't quite describe it precisely, but I'll give it a shot anyway. Because Scala allows "mix in" of traits, there is, CMIIW, no hidden container-managed proxy code is needed. Let's compare this with CDI used in the JSF framework. To cater to different scopes, a CDI bean can be annotated with @RequestScoped, @SessionScoped, @ConversationScoped, etc. How does it look like when constructed? Honestly I don't know. I only know how it should work thanks to the CDI spec's documentation. But I think the container would have to create a specific proxy for each bean depending on its scope, and inject its proxy dependencies and so on. The proxy might also be different when you add @Stateful to, say, a @ConversationScoped bean.

On the other hand, in Lift, it's much simpler. Snippets are normal methods registered via the LiftRules object during boot. Request and session states are just plain objects extending either RequestVar or SessionVar, used directly by snippets. The separation between controller logic and session/request data is clear here. You know what's happening to your stuff by checking out what's inside what you're extending.