-
Notifications
You must be signed in to change notification settings - Fork 267
The producer api
TODO: Intro to the Producer API.
You get data into Summingbird by instantiating a Source
, defined by the given execution platform, and passing it to Producer.source
:
def source[P <: Platform[P], T: Manifest](s: P#Source[T]): Producer[P, T]
Producer.source
returns a leaf node in the graph. Calling methods on the resulting producer won’t actually do anything! Summingbird’s DSL is declarative. You’re building up a computation that will only execute later, when you use some particular Platform to plan the topology.
The map-like functions available to a producer operate on individual elements in the stream represented by the producer. In each of the examples below, producer
is of type Producer[P <: Platform[P], T]
and represents a stream of T
using whatever representation is required by the abstract execution platform P
.
The map
method takes a function from T => U
and performs the transformation on each item in the stream. The result is a Producer[P, U]
.
val birds: Producer[P, Bird] = ???
val speeds: Producer[P, Int] =
birds.map { bird: Bird => bird.getSpeed }
The filter
method takes a function from T => Boolean
and filters out all items from the stream for which the predicate returns false
. The result is a Producer[P, T]
; a stream of the same type, with potentially fewer items.
val quickBirds: Producer[P, Bird] =
birds.filter { bird: Bird => bird.getSpeed > MilesPerHour(10) }
flatMap
maps each T
in the producer’s stream to a TraversableOnce[U]
and then flattens together all returned lists. The return value of flatMap
is a Producer[P, U]
.
val chicks: Producer[P, Chick] =
birds.flatMap { bird: Bird => bird.getChildren: Seq[Chick] }
The producer returned by the flatMap
function may have the same, greater or fewer items than the original stream.
Calling name(identifier)
on a producer assigns a name to every node above the call. Platform implementations use this feature to allow the user to pass in options at planning time that are meant for sections of a given producer graph. For example:
birds.name("preOperations")
.filter { bird: Bird => bird.getSpeed > MilesPerHour(10) }
.name("filtered")
.flatMap { bird: Bird => bird.getChildren: Seq[Chick] }
.name("flatMapped")
The original birds
branch is now named “preOperations”. The filter
call occurs in a branch named “filtered”, and the flatMap
call occurs in a branch named flatMapped
.
TODO. Merge combines two streams without any uniqueing or joining.
TODO. either
is a way to merge with a producer of another type.
leftJoin takes a service and performs a left join, or a lookup join, in some way that makes sense to the underlying platform. In realtime, this looks like a key-value lookup on a random access key value store. In offline mode, we do a special kind of join that simulates the kv pair lookup at a specific time. More information on that later, after we describe how to write out streams of data.
def leftJoin[RightV](service: P#Service[K, RightV]): KeyedProducer[P, K, (V, Option[RightV])] =
There are two ways to get data out of the system. You can write into a Sink, or you can sumByKey into a Store. The sink accepts a stream of data; each instance is associated with a time. In realtime, this just looks like a queue. Another job might consume this.
write
writes out a stream of the data.
sumByKey
allows you to write out a snapshot of data.
The CacheSize
option controls Summingbird’s map-side aggregation. If you apply a CacheSize(n)
option on a flatMap phase, Summingbird will buffer n
key-value pairs on each flatMap bolt executor before it does an emit. For example:
val tail = source.flatMap { t => ... }.name("my FM")
val s = Storm.local(Map("my FM" -> Options.set(CacheSize(100))))
s.plan(tail)
In this scenario, each of the 20 executors will buffer 100 key-value pairs. When each executor hits 100, it’ll sum all pairs up by key and emit the aggregated values. This is Summingbird’s realtime version of map-side aggregation, and can be really useful when your key space isn’t that large, or when you have huge skew in your data (since the skewed key will be pre-aggregated instead of kicking out large numbers of values).
If you apply a CacheSize
option to a sumByKey
, Summingbird will do the same sort of buffering in the SummerBolt
on its commit out to the underlying Store
. Same exact idea as the flatMap
cache size, just another knob that you get to control.