blob: 555d2aed5602e22f7271bc2e34e658cb2e4a14fe [file] [view]
---
layout: global
title: Extending Spark with Spark Server Libraries
redirect_from: app-dev-spark-connect.html
license: |
Licensed to the Apache Software Foundation (ASF) under one or more
contributor license agreements. See the NOTICE file distributed with
this work for additional information regarding copyright ownership.
The ASF licenses this file to You under the Apache License, Version 2.0
(the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
---
With its decoupled client-server architecture, Spark Connect simplifies how Spark Applications are
developed.
The notion of Spark Client Applications and Spark Server Libraries are introduced as follows:
* _Spark Client Applications_ are regular Spark applications that use Spark and its rich ecosystem for
distributed data processing. Examples include ETL pipelines, data preparation, and model training
and inference.
* _Spark Server Libraries_ build on, extend, and complement Spark's functionality, e.g.
[MLlib](ml-guide.html) (distributed ML libraries that use Spark's powerful distributed processing). Spark Connect
can be extended to expose client-side interfaces for Spark Server Libraries.
With Spark 3.4 and Spark Connect, the development of Spark Client Applications is simplified, and
clear extension points and guidelines are provided on how to build Spark Server Libraries, making
it easy for both types of applications to evolve alongside Spark. As illustrated in Fig.1, Spark
Client applications connect to Spark using the Spark Connect API, which is essentially the
DataFrame API and fully declarative.
<figure style="text-align: center;">
<img
src="img/extending-spark-connect-labelled.png"
alt="Extending Spark Connect Diagram"
/>
<figcaption>Figure 1: Architecture</figcaption>
</figure>
Spark Server Libraries extend Spark. They typically provide additional server-side logic integrated
with Spark, which is exposed to client applications as part of the Spark Connect API, using Spark
Connect extension points. For example, the _Spark Server Library_ consists of custom
server-side logic (the blue box labeled _Plugin_, with its _App Logic_), which is exposed to
the client via the blue _Proto API_ box as part of the Spark Connect API. The client uses this
API, e.g., alongside PySpark or the Spark Scala client, making it easy for Spark client
applications to work with the custom logic/library.
Until Spark 3.4, extensions to Spark (e.g., [Spark ML](ml-guide#:~:text=What%20is%20%E2%80%9CSpark%20ML%E2%80%9D%3F,to%20emphasize%20the%20pipeline%20concept.)
or [Spark-NLP](https://github.com/JohnSnowLabs/spark-nlp)) were built and deployed like Spark
Client Applications. With Spark 3.4 and Spark Connect, explicit extension points are offered to
extend Spark via Spark Server Libraries. These extension points provide functionality that can be
exposed to a client, which differs from existing extension points in Spark such as
[SparkSession extensions](api/java/org/apache/spark/sql/SparkSessionExtensions.html) or
[Spark Plugins](api/java/org/apache/spark/api/plugin/SparkPlugin.html).
* This will become a table of contents (this text will be scraped).
{:toc}
A Spark Server Library consists of the following components, numbered in Fig. 1:
1. The Spark Connect protocol extension (blue box _Proto_ API)
2. A Spark Connect Plugin.
3. The application logic that extends Spark.
4. The client package that exposes the Spark Server Library application logic to the Spark Client
Application, alongside PySpark or the Scala Spark Client.
## Spark Connect Protocol Extension
To extend Spark with a new Spark Server Library, developers can extend the three main operation
types in the Spark Connect protocol: _Relation_, _Expression_, and _Command_.
```protobuf
message Relation {
oneof rel_type {
Read read = 1;
// ...
google.protobuf.Any extension = 998;
}
}
message Expression {
oneof expr_type {
Literal literal = 1;
// ...
google.protobuf.Any extension = 999;
}
}
message Command {
oneof command_type {
WriteCommand write_command = 1;
// ...
google.protobuf.Any extension = 999;
}
}
```
Their extension fields allow serializing arbitrary protobuf messages as part of the Spark Connect
protocol. These messages represent the parameters or state of the extension implementation.
To build a custom expression type, the developer first defines the custom protobuf definition
of the expression.
```protobuf
message ExamplePluginExpression {
Expression child = 1;
string custom_field = 2;
}
```
## Spark Connect Plugin implementation with custom application logic
As a next step, the developer implements the `ExpressionPlugin` class of Spark Connect with custom
application logic based on the input parameters of the protobuf message.
```scala
class ExampleExpressionPlugin extends ExpressionPlugin {
override def transform(
relation: protobuf.Any,
planner: SparkConnectPlanner): Option[Expression] = {
// Check if the serialized value of protobuf.Any matches the type
// of our example expression.
if (!relation.is(classOf[proto.ExamplePluginExpression])) {
return None
}
val exp = relation.unpack(classOf[proto.ExamplePluginExpression])
Some(Alias(planner.transformExpression(
exp.getChild), exp.getCustomField)(explicitMetadata = None))
}
}
```
Once the application logic is developed, the code must be packaged as a jar and Spark must be
configured to pick up the additional logic. The relevant Spark configuration options are:
* `spark.jars` which define the location of the Jar file containing the application logic built for
the custom expression.
* `spark.connect.extensions.expression.classes` specifying the full class name
of each expression extension loaded by Spark. Based on these configuration options, Spark will
load the values at startup and make them available for processing.
## Spark Server Library Client Package
Once the server component is deployed, any client can use it with the right protobuf messages.
In the example above, the following message payload sent to the Spark Connect endpoint would be
enough to trigger the extension mechanism.
```json
{
"project": {
"input": {
"sql": {
"query": "select * from samples.nyctaxi.trips"
}
},
"expressions": [
{
"extension": {
"typeUrl": "type.googleapis.com/spark.connect.ExamplePluginExpression",
"value": "\n\006\022\004\n\002id\022\006testval"
}
}
]
}
}
```
To make the example available in Python, the application developer provides a Python library that
wraps the new expression and embeds it into PySpark. The easiest way to provide a function for any
expression is to take a PySpark column instance as an argument and return a new Column instance
with the expression applied.
```python
from pyspark.sql.connect.column import Expression
import pyspark.sql.connect.proto as proto
from myxample.proto import ExamplePluginExpression
# Internal class that satisfies the interface by the Python client
# of Spark Connect to generate the protobuf representation from
# an instance of the expression.
class ExampleExpression(Expression):
def to_plan(self, session) -> proto.Expression:
fun = proto.Expression()
plugin = ExamplePluginExpression()
plugin.child.literal.long = 10
plugin.custom_field = "example"
fun.extension.Pack(plugin)
return fun
# Defining the function to be used from the consumers.
def example_expression(col: Column) -> Column:
return Column(ExampleExpression())
# Using the expression in the Spark Connect client code.
df = spark.read.table("samples.nyctaxi.trips")
df.select(example_expression(df["fare_amount"])).collect()
```