Tag Archives: postgresql

ETL & Pipelining With Alpakka Kafka

With the increasing demand for big data transformations on distributed platforms, one approach is to put in place a streaming ETL system with built-in packpressure using Alpakka Kafka API that allows composable pipelines to be constructed on-demand. This blog post extends the previous post which shows a skeletal version of the system.

Let’s first review the diagram shown in the previous post of what we’re aiming to build — a streaming ETL system empowered by reactive streams and Apache Kafka’s publish-subscribe machinery for durable stream data to be produced or consumed by various data processing/storage systems:

Alpakka Kafka - Streaming ETL

In this blog post, we’re going to:

  1. enhance the data warehouse consumer to programmatically track the commit offset positions,
  2. plug into an existing consumer a data processing pipeline as a stream processing operator, and,
  3. add to the streaming ETL system a mix of heterogeneous producers and consumers

Action item #1 would address the requirement of at-least-once delivery in stream consumption. #2 illustrates how to add to the streaming ETL system a custom data pipeline as a composable stream flow, and #3 showcases how data in various storage systems can participate in the real-time stream to operate (serially or in parallel) as composable sources, flows and sinks. All relevant source code is available in this GitHub repo.

Real-time streaming ETL/pipelining of property listing data

For usage demonstration, the application runs ETL/pipelining of data with a minified real estate property listing data model. It should be noted that expanding or even changing it altogether to a different data model should not affect how the core streaming ETL system operates.

Below are a couple of links related to library dependencies and configurations for the core application:

  • Library dependencies in build.sbt [separate tab]
  • Configurations for Akka, Kafka, PostgreSQL & Cassandra in application.conf [separate tab]

It’s also worth noting that the application can be scaled up with just configurative changes. For example, if the Kafka brokers and Cassandra database span across multiple hosts, relevant configurations like Kafka’s bootstrap.servers could be "10.1.0.1:9092,10.1.0.2:9092,10.1.0.3:9092" and contact-points for Cassandra might look like ["10.2.0.1:9042","10.2.0.2:9042"].

Next, let’s get ourselves familiarized with the property listing data definitions in the PostgreSQL and Cassandra, as well as the property listing classes that model the schemas.

A Kafka producer using Alpakka Csv

Alpakka comes with a simple API for CSV file parsing with method lineScanner() that takes parameters including the delimiter character and returns a Flow[ByteString, List[ByteString], NotUsed].

Below is the relevant code in CsvPlain.scala that highlights how the CSV file gets parsed and materialized into a stream of Map[String,String] via CsvParsing and CsvToMap, followed by transforming into a stream of PropertyListing objects.

Note that the drop(offset)/take(limit) code line, which can be useful for testing, is for taking a segmented range of the stream source and can be removed if preferred.

A subsequent map wraps each of the PropertyListing objects in a ProducerRecord[K,V] with the associated topic and key/value of type String/JSON before being streamed into Kafka via Alpakka Kafka’s Producer.plainSink().

A Kafka producer using Alpakka Slick

The PostgresPlain producer, which is pretty much identical to the one described in the previous blog post, creates a Kafka producer using Alpakka Slick which allows SQL queries into a PostgreSQL database to be coded in Slick’s functional programming style.

The partial code below shows how method Slick.source() takes a streaming query and returns a stream source of PropertyListing objects.

The high-level code logic in PostgresPlain is similar to that of the CsvPlain producer.

A Kafka consumer using Alpakka Cassandra

We created a Kafka consumer in the previous blog post using Alpakka Kafka’s Consumer.plainSource[K,V] for consuming data from a given Kafka topic into a Cassandra database.

The following partial code from the slightly refactored version of the consumer, CassandraPlain shows how data associated with a given Kafka topic can be consumed via Alpakka Kafka’s Consumer.plainSource().

Alpakka’s CassandraFlow.create() is the stream processing operator responsible for funneling data into the Cassandra database. Note that it takes a CQL PreparedStatement along with a “statement binder” that binds the incoming class variables to the corresponding Cassandra table columns before executing the CQL.

Enhancing the Kafka consumer for ‘at-least-once’ consumption

To enable at-least-once consumption by Cassandra, instead of Consumer.plainSource[K,V], we construct the stream graph via Alpakka Kafka Consumer.committableSource[K,V] which offers programmatic tracking of the commit offset positions. By keeping the commit offsets as an integral part of the streaming data, failed streams could be re-run from the offset positions.

The main stream composition code of the enhanced consumer, CassandraCommittable.scala, is shown below.

A couple of notes:

  1. In order to be able to programmatically keep track of the commit offset positions, each of the stream elements emitted from Consumer.committableSource[K,V] is wrapped in a CommittableMessage[K,V] object, consisting of the CommittableOffset value in addition to the Kafka ConsumerRecord[K,V].
  2. Committing the offset should be done after the stream data is processed for at-least-once consumption, whereas committing prior to processing the stream data would only achieve at-most-once delivery.

Adding a property-rating pipeline to the Alpakka Kafka consumer

Next, we add a data processing pipeline to the consumer to perform a number of ratings of the individual property listings in the stream before delivering the rated property listing data to the Cassandra database, as illustrated in the following diagram.

Alpakka Kafka - Streaming ETL w/ custom pipelines

Since the CassandraFlow.create() stream operator will be executed after the rating pipeline, the corresponding “statement binder” necessary for class-variable/table-column binding will now need to encapsulate also PropertyRating along with CommittableMessage[K,V], as shown in the partial code of CassandraCommittableWithRatings.scala below.

For demonstration purpose, we create a dummy pipeline for rating of individual real estate properties in areas such as affordability, neighborhood, each returning just a Future of random integers between 1 and 5 after a random time delay. The rating related fields along with the computation logic are wrapped in class PropertyRating as shown below.

A Kafka consumer with a custom flow & stream destination

The application is also bundled with a consumer with the property rating pipeline followed by a custom flow to showcase how one can compose an arbitrary side-effecting operator with custom stream destination.

Note that mapAsync is used to allow the stream transformation by the custom business logic to be carried out asynchronously.

Running the streaming ETL/pipelining system

To run the application that comes with sample real estate property listing data on a computer, go to the GitHub repo and follow the README instructions to launch the producers and consumers on one or more command-line terminals.

Also included in the README are instructions about how to run a couple of set queries to verify data that get ETL-ed to the Cassandra tables via Alpakka Cassandra’s CassandraSource which takes a CQL query as its argument.

Further enhancements

Depending on specific business requirement, the streaming ETL system can be further enhanced in a number of areas.

  1. This streaming ETL system offers at-least-once delivery only in stream consumptions. If an end-to-end version is necessary, one could enhance the producers by using Producer.committabbleSink() or Producer.flexiFlow() instead of Producer.plainSink().
  2. For exactly-once delivery, which is a generally much more stringent requirement, one approach to achieve that would be to atomically persist the in-flight data with the corresponding commit offset positions using a reliable storage system.
  3. In case tracking of Kafka’s topic partition assignment is required, one can use Consumer.committablePartitionedSource[K,V] instead of Consumer.committableSource[K,V]. More details can be found in the tech doc.
  4. To gracefully restart a stream on failure with a configurable backoff, Akka Stream provides method RestartSource.onFailuresWithBackoff for that as illustrated in an example in this tech doc.

Streaming ETL With Alpakka Kafka

In a previous startup I cofounded, our core product was a geospatial application that provided algorithmic ratings of the individual residential real estate properties for home buyers. Given that there were over 100+ millions of residential properties nationwide, the collective data volume of all the associated attributes necessary for the data engineering work was massive.

For the initial MVP (minimum viable product) releases in which we only needed to showcase our product features in a selected metropolitan area, we used PostgreSQL as the OLTP (online transaction processing) database. Leveraging Postgres’ table partitioning feature, we had an OLTP database capable of accommodating incremental geographical expansion into multiple cities and states.

Batch ETL

The need for a big data warehouse wasn’t imminent in the beginning, though we had to make sure a data processing platform for a highly scalable data warehouse along with efficient ETL (extract/transform/load) functions would be ready on a short notice. The main objective was to make sure the OLTP database could be kept at a minimal volume while less frequently used data got “archived” off to a big data warehouse for data analytics.

With limited engineering resources available in a small startup, I kicked off a R&D project on the side to build programmatic ETL processes to periodically funnel data from PostgreSQL to a big data warehouse in a batch manner. Cassandra was chosen to be the data warehouse and was configured on an Amazon EC2 cluster. The project was finished with a batch ETL solution that functionally worked as intended, although back in my mind a more “continuous” operational model would be preferred.

Real-time Streaming ETL

Fast-forward to 2021, I recently took on a big data streaming project that involves ETL and building data pipelines on a distributed platform. Central to the project requirement is real-time (or more precisely, near real-time) processing of high-volume data. Another aspect of the requirement is that the streaming system has to accommodate custom data pipelines as composable components of the consumers, suggesting that a streaming ETL solution would be more suitable than a batch one. Lastly, stream consumption needs to guarantee at-least-once delivery.

Given all that, Apache Kafka promptly stood out as a top candidate to serve as the distributed streaming brokers. In particular, its capability of keeping durable data in a distributed fault-tolerant cluster allows it to serve different consumers at various instances of time and locales. Next, Akka Stream was added to the tech stack for its versatile stream-based application integration functionality as well as benefits of reactive streams.

Alpakka – a reactive stream API and DSL

Built on top of Akka Stream, Alpakka provides a comprehensive API and DSL (domain specific language) for reactive and stream-oriented programming to address the application integration needs for interoperating with a wide range of prominent systems across various computing domains. That, coupled with the underlying Akka Stream’s versatile streaming functions, makes Alpakka a powerful toolkit for what is needed.

In this blog post, we’ll assemble in Scala a producer and a consumer using the Alpakka API to perform streaming ETL from a PostgreSQL database through Kafka brokers into a Cassandra data warehouse. In a subsequent post, we’ll enhance and package up these snippets to address the requirement of at-least-once delivery in consumption and composability of data pipelines.

Streaming ETL with Alpakka Kafka, Slick, Cassandra, …

The following diagram shows the near-real time ETL functional flow of data streaming from various kinds of data sources (e.g. a PostgreSQL database or a CSV file) to data destinations (e.g. a Cassandra data warehouse or a custom data stream outlet).

Alpakka Kafka - Streaming ETL

The Apache Kafka brokers provide a distributed publish-subscribe platform for keeping in-flight data in durable immutable logs readily available for consumption. Meanwhile, the Akka Stream based Alpakka API that comes with a DSL allows programmatic integrations to compose data pipelines as sources, sinks and flows, in addition to enabling “reactivity” by equipping the streams with non-blocking backpressure.

It should be noted that the same stream can be processed using various data sources and destinations simultaneously. For instance, data with the same schema from both the CSV file and Postgres database could be published to the same topic and consumed by a consumer group designated for the Cassandra database and another consumer group for a different data storage.

Example: ETL of real estate property listing data

The platform will be for general-purpose ETL/pipelining. For illustration purpose in this blog post, we’re going to use it to perform streaming ETL of some simplified dataset of residential real estate property listings.

First, we create a simple class to represent a property listing.

Using the good old sbt as the build tool, relevant library dependencies for Akka Stream, Alpakka Kafka, Postgres/Slick and Cassandra/DataStax are included in build.sbt.

Next, we put configurations for Akka Actor, Alpakka Kafka, Slick and Cassandra in application.conf under src/main/resources/:

Note that the sample configuration is for running the application with all Kafka, PostgreSQL and Cassandra on a single computer. The host IPs (i.e. 127.0.0.1) should be replaced with their corresponding host IPs/names in case they’re on separate hosts. For example, relevant configurations for Kafka brokers and Cassandra database spanning across multiple hosts might look something like bootstrap.servers = "10.1.0.1:9092,10.1.0.2:9092,10.1.0.3:9092" and contact-points = ["10.2.0.1:9042","10.2.0.2:9042"].

PostgresProducerPlain – an Alpakka Kafka producer

The PostgresProducerPlain snippet below creates a Kafka producer using Alpakka Slick which allows SQL queries to be coded in Slick’s functional programming style.

Method Slick.source[T]() takes a streaming query and returns a Source[T, NotUsed]. In this case, T is PropertyListing. Note that Slick.source() can also take a plain SQL statement wrapped within sql"..." as its argument, if wanted (in which case an implicit value of slick.jdbc.GetResult should be defined).

A subsequent map wraps each of the property listing objects in a ProducerRecord[K,V] with topic and key/value of type String/JSON, before publishing to the Kafka topic via Alpakka Kafka’s Producer.plainSink[K,V].

To run PostgresProducerPlain, simply navigate to the project root and execute the following command from within a command line terminal:

CassandraConsumerPlain – an Alpakka Kafka consumer

Using Alpakka Kafka, CassandraConsumerPlain shows how a basic Kafka consumer can be formulated as an Akka stream that consumes data from Kafka via Consumer.plainSource followed by a stream processing operator, Alpakka Cassandra’s CassandraFlow to stream the data into a Cassandra database.

A few notes:

  • Consumer.plainSource: As a first stab at building a consumer, we use Alpakka Kafka’s Consumer.plainSource[K,V] as the stream source. To ensure the stream to be stopped in a controlled fashion, Consumer.Drainingcontrol is included when composing the stream graph. While it’s straight forward to use, plainSource doesn’t offer programmatic tracking of the commit offset position thus cannot guarantee at-least-once delivery. An enhanced version of the consumer will be constructed in a subsequent blog post.
  • Partition key: Cassandra mandates having a partition key as part of the primary key of every table for distributing across cluster nodes. In our property listing data, we make the modulo of the Postgres primary key property_id by the number of partitions to be the partition key. It could certainly be redefined to something else (e.g. locale or type of the property) in accordance with the specific business requirement.
  • CassandraSource: Method query() simply executes queries against a Cassandra database using CassandraSource which takes a CQL query with syntax similar to standard SQL’s. It isn’t part of the consumer flow, but is rather as a convenient tool for verifying stream consumption result.
  • CassandraFlow: Alpakka’s CassandraFlow.create[S]() is the main processing operator responsible for streaming data into the Cassandra database. It takes a CQL PreparedStatement and a “statement binder” that binds the incoming class variables to the corresponding Cassandra columns before executing the insert/update. In this case, S is ConsumerRecord[K,V].

To run CassandraConsumerPlain, Navigate to the project root and execute the following from within a command line terminal:

Table schema in PostgreSQL & Cassandra

Obviously, the streaming ETL application is supposed to run in the presence of one or more Kafka brokers, a PostgreSQL database and a Cassandra data warehouse. For proof of concept, getting all these systems with basic configurations on a descent computer (Linux, Mac OS, etc) is a trivial exercise. The ETL application is readily scalable that it would require only configurative changes when, say, needs rise for scaling up of Kafka and Cassandra to span clusters of nodes in the cloud.

Below is how the table schema of property_listing can be created in PostgreSQL via psql:

To create keyspace propertydata and the corresponding table property_listing in Cassandra, one can launch cqlsh and execute the following CQL statements:

What’s next?

So, we now have a basic streaming ETL system running Alphakka Kafka on top of a cluster of Kafka brokers to form the reactive stream “backbone” for near real-time ETL between data stores. With Alpakka Slick and Alpakka Cassandra, a relational database like PostgreSQL and a Cassandra data warehouse can be made part of the system like composable stream components.

As noted earlier, the existing Cassandra consumer does not guarantee at-least-once delivery, which is part of the requirement. In the next blog post, we’ll enhance the existing consumer to address the required delivery guarantee. We’ll also add a data processing pipeline to illustrate how to construct additional data pipelines as composable stream operators. All relevant source code along with some sample dataset will be published in a GitHub repo.

PostgreSQL Table Partitioning

With the ever growing demand for data science work in recent years, PostgreSQL has gained superb popularity especially in areas where extensive geospatial/GIS (geographic information system) functionality is needed. In a previous startup venture, MySQL was initially adopted and I went through the trouble of migrating to PostgreSQL mainly because of the sophisticated geospatial features PostGIS offers.

PostgreSQL offers a lot of goodies, although it does have a few things that I wish were done differently. Most notable to me is that while its SELECT statement supports SQL-92 Standard’s JOIN syntax, its UPDATE statement would not. For instance, the following UPDATE statement would not work in PostgreSQL:

Partial indexing

Nevertheless, for general performance and scalability, PostgreSQL remains one of the top candidates with proven track record in the world of open source RDBMS. In scaling up a PostgreSQL database, there is a wide variety of approaches. Suitable indexing is probably one of the first things to look into. Aside from planning out proper column orders in indexes that are optimal for the frequently used queries, there is another indexing feature that PostgreSQL provides for handling large datasets.

Partial indexing allows an index to be built over a subset of a table based on a conditional expression. For instance:

In the case of a table with large amount of rows, this feature could make an otherwise gigantic index much smaller, thus more efficient for queries against the selectively indexed data.

Scaling up with table partitioning

However, when a table grows to certain volume, say, beyond a couple of hundreds of million rows, and if periodically archiving off data from the table isn’t an option, it would still be a problem even with applicable indexing strategy. In many cases, it might be necessary to do something directly with the table structure and table partitioning is often a good solution.

There are a few approaches to partition a PostgreSQL table. Among them, partitioning by means of table inheritance is perhaps the most popular approach. A master table will be created as a template that defines the table structure. This master table will be empty whereas a number of child tables inherited from this master table will actually host the data.

The partitioning is based on a partition key which can be a column or a combination of columns. In some common use cases, the partition keys are often date-time related. For instance, a partition key could be defined in a table to partition all sales orders by months with constraint like the following:

order_date >= ‘2016-12-01 00:00:00’ AND order_date < ‘2017-01-01 00:00:00’

Other common cases include partitioning geographically, etc.

A table partitioning example

When I was with a real estate startup building an application that involves over 100 millions nationwide properties, each with multiple attributes of interest, table partitioning was employed to address the demanding data volume. Below is a simplified example of how the property sale transaction table was partitioned to maintain a billion rows of data.

First, create the master table which will serve as the template for the table structure.

Next, create child tables inheriting from the master table for the individual states. For simplicity, I only set up 24 states for performance evaluation.

Nothing magical so far, until a suitable trigger for propagating insert is put in place. The trigger essentially redirects insert requests against the master table to the corresponding child tables.

Let’s test inserting data into the partitioned tables via the trigger:

A Python program for data import

Now that the master table and its child tables are functionally in place, we’re going to populate them with large-scale data for testing. First, write a simple program using Python (or any other programming/scripting language) as follows to generate simulated data in a tab-delimited file for data import:

Run the Python program to generate up to 1 billion rows of property sale data. Given the rather huge output, make sure the generated file is on a storage device with plenty of space. Since it’s going to take some time to finish the task, it would better be run in the background, perhaps along with mail notification, like the following:

Next, load data from the generated infile into the partitioned tables using psql. In case there are indexes created for the partitioned tables, it would generally be much more efficient to first drop them and recreate them after loading the data, like in the following:

Query with Constraint Exclusion

Prior to querying the tables, make sure the query optimization parameter, constraint_exclusion, is enabled.

With constraint exclusion enabled, the query planner will be smart enough to examine query constraints to exclude scanning of those partitioned tables that don’t match the constraints. Unfortunately, though, if the constraints involve matching against non-constants like the NOW() function, the query planner won’t have enough information to filter out unwanted partitions hence won’t be able to take advantage of the optimization.

Final notes

With a suitable partitioning scheme applied to a big table, query performance can be improved by an order of magnitude. As illustrated in the above case, the entire partitioning scheme centers around the key column used for partitioning, hence it’s critical to properly plan out which key column (or combination of columns) to partition. Number of partitions should also be carefully thought out, as too few partitions might not help whereas too many partitions would create too much overhead.