Lazy Instantiation of Dataset from Amazon S3

Target audience: Intermediate
Estimated reading time: 4'

Posts history

Have you ever wished for the capability to instantaneously instantiate an object or dataset stored on Amazon S3 as the need arises? With Scala and Spark's functional toolkit, this is more straightforward than it appears.

The methodology for lazy instantiation of datasets from S3 was initially crafted using Scala 2.12 and Apache Spark 2.4. Notably, this code isn't tied to a particular version of the language or framework and remains compatible with Apache Spark 3.0.

Table of contents

Just be lazy

A simple, efficient repository

Data loader

References

Just be lazy

A common requirement in machine learning is to load the configuration parameters associated with model at run-time. The model may have been trained with data segregated by customers, or categories. When deployed for prediction, it is critical to select/load the right set of parameters according to the characteristic of the request
For instance, a topic extraction model may have been trained with scientific corpus, medical articles or computer science papers. 

A simple approach is to pre-load all variants of a model when the underlying application is deployed in production. However, consuming uncessary memory and CPU cycle for a model that may be needed, at least right away, is a waste of resource. In this post, we assume that the model parameters are stored on Amazon S3.
Lazy instantiation of objects allows us to reduce unnecessary memory consumption by invoking a constructor once, only when needed. This capability becomes critical for data with large footprint such as Apache Spark data sets.

A simple, efficient repository

Let's consider all the credentials to access multiple devices consisting of an id, password and hint that have been previously uploaded on S3. 

case class Credentials(

   device: String, 

   id: String, 

   password: String, 

   hint: String

)

A hash table is the simplest incarnation of a dynamic repository of models. Therefore we implement a lazy hash table by sub-classing the mutable HashMap.
The first time a model is requested, it is loaded into memory from S3 that returned to the client code. To this purpose we need to define the following argument for the constructor of the lazy hash table

• Dynamic loading mechanism from S3 - loader is responsible for loading the data from S3

• Key generator - toKey converts a string key to a the type of key of the Hash map

final class LazyHashMap[T, U](
     loader: String => Option[U], 
     toKey: String => T) extends HashMap[T, U] {

     // Override the HashMap.get method 
   override def get(item: String): Option[U] = synchronized {
       val key = toKey(item)

       if(super.contains(key)) // Is is already in memory?
           super.get(key) 

      else
           loader(item).map(  // otherwise load the item from S3
              l => {
                super.put(key, l)
                l
             }
          )
    }

      // Prevent for updating this immutable map

   @throws(class = classOf[UnsupportedOperationException])

   override def put(key: T, value: U): Option[U] 

         throw new UnsupportedOperationException("lazy map is immutable")

The keyword synchronized implements a critical section to protect the execution from dirty read. 
Here is an example of the two arguments for the constructor of the lazy hash table for a type MyValue. The key identifies the data set and the model which has been trained on.

val load: String => Option[Credentials] = 

     (dataSource: String) => loadData(dataSource)
val key = (s: String) => s

val lazyHashMap = new LazyHashMap[String, Dataset[Credentials]](load, key)

The last business to take care of is the implementation of the function, loadData to load and instantiate the dataset

Data loader

Let's write a loader for a Spark data set of type T stored on AWS S3 in a given bucket, bucketName and folder, s3InputPath

def s3ToDataset[T](
     s3InputPath: String
)(implicit encoder: Encoder[T]): Dataset[T] = {
   import sparkSession.implicits._

    // Needed for access keys and infer schema
   val loadDS = Seq[T]().toDS
   val accessConfig = loadDS.sparkSession 

         .sparkContext 

         .hadoopConfiguration

   // Credentials to read from S3
   accessConfig.set("fs.s3a.access.key", myAccessKey)
   accessConfig.set("fs.s3a.secret.key", mySecretKey)

   try {

       // Enforce the schema
      val inputSchema = loadDS.schema
      sparkSession.read
	 .format("json")
	 .schema(inputSchema)
	 .load(path = s"s3a://$bucketName/${s3InputPath}")
	 .as[T]
   }
   catch {
      case e: FileNotFoundException => log.error(e.getMessage)
      case e: SparkException => log.error(e.getMessage)
      case e: IOException =>  log.error(e.getMessage)
   }
}

It is assumed that the Apache Spark session has already been created and an encoder (i.e. Kryo) has been already been defined. The encoder for the type T is implicitly defined, usually along with the Spark session.

The first step is to instantiate a 'dummy' empty dataset of type T. The instantiation, loadDS is used to

• Access the Hadoop configuration to specify the credentials for S3
• Enforce the schema when reading the data (in JSON) format from S3. Alternatively, the schema could have been inferred.

Note Data from S3 bucket is accessed through the s3a:// protocol. It add an object layer on top of the default S3 protocol which is block-centric. It is significantly faster.

Finally let's implement the load function, loadData

  // Create a simple Spark session

implicit val sparkSession =  SparkSession.builder
     .appName("ExecutionContext").config(conf)
     .getOrCreate()

def loadData(s3Path: String): Option[Dataset[Credentials]] = {
    import sparkSession.implicits._ // need for encoding
    s3ToDataset[Credentials]
}

Thank you for reading this article. For more information ...

References

• Comparing s3a, s3n and s3 protocol
• Apache Spark quick start
• github.com/patnicolas

---------------------------

Patrick Nicolas has over 25 years of experience in software and data engineering, architecture design and end-to-end deployment and support with extensive knowledge in machine learning. 
He has been director of data engineering at Aideo Technologies since 2017 and he is the author of "Scala for Machine Learning" Packt Publishing ISBN 978-1-78712-238-3