Kafka Connect FTP

Monitors files on an FTP server and feeds changes into Kafka.

Remote directories of interest are to be provided. On a specified interval, the list of files in the directories is refreshed. Files are downloaded when they were not known before, or when their timestamp or size are changed. Only files with a timestamp younger than the specified maximum age are considered. Hashes of the files are maintained and used to check for content changes. Changed files are then fed into Kafka, either as a whole (update) or only the appended part (tail), depending on the configuration. Optionally, file bodies can be transformed through a pluggable system prior to putting it into Kafka.

Data Types

Each Kafka record represents a file, and has the following types.

• The format of the keys is configurable through ftp.keystyle=string|struct. It can be a string with the file name, or a FileInfo structure with name: string and offset: long. The offset is always 0 for files that are updated as a whole, and hence only relevant for tailed files.
• The values of the records contain the body of the file as bytes.

Setup

Properties

In addition to the general configuration for Kafka connectors (e.g. name, connector.class, etc.) the following options are available.

An example file is here.

Tailing Versus Update as a Whole

The following rules are used.

• Tailed files are only allowed to grow. Bytes that have been appended to it since a last inspection are yielded. Preceding bytes are not allowed to change;
• Updated files can grow, shrink and change anywhere. The entire contents are yielded.

Usage

Build.

mvn clean package

Put jar into CLASSPATH.

export CLASSPATH=`realpath ./target/kafka-connect-ftp-0.1-jar-with-dependencies.jar` 

With $CONFLUENT_HOME pointing to the root of your Confluent Platform installation, start.

$CONFLUENT_HOME/bin/connect-standalone $CONFLUENT_HOME/etc/schema-registry/connect-avro-standalone.properties your.specific.properties

Data Converters

Instead of dumping whole file bodies (and the danger of exceeding Kafka's message.max.bytes), one might want to give an interpretation to the data contained in the files before putting it into Kafka. For example, if the files that are fetched from the FTP are comma-separated values (CSVs), one might prefer to have a stream of CSV records instead. To allow to do so, the connector provides a pluggable conversion of SourceRecords. Right before sending a SourceRecord to the Connect framework, it is run through an object that implements:

package com.eneco.trading.kafka.connect.ftp.source

trait SourceRecordConverter extends Configurable {
  def convert(in:SourceRecord) : java.util.List[SourceRecord]
}

(for the Java people, read: interface instead of trait).

The default object that is used is a pass-through converter, an instance of:

class NopSourceRecordConverter extends SourceRecordConverter{
  override def configure(props: util.Map[String, _]): Unit = {}
  override def convert(in: SourceRecord): util.List[SourceRecord] = Seq(in).asJava
}

To override it, create your own implementation of SourceRecordConverter, put the jar into your $CLASSPATH and instruct the connector to use it via the .properties:

ftp.sourcerecordconverter=your.name.space.YourConverter