
Sign up to save your podcasts
Or


At Airbnb, infrastructure management is standardized across the organization. Platform engineering teams build tools that allow the other teams throughout the organization to work more effectively. A platform engineering team handles problems such as continuous integration, observability, and service discovery.
Other teams throughout a company use the tools that a platform engineering team builds. For example, there is a team at Airbnb that builds the search and discovery system that is used by customers who are looking for a place to stay. That team does not want to have to worry about how they are deploying, how their service is being logged, and how to scale up. All of that should be taken care of by the platform engineering team.
At a large company like Airbnb, there is so much happening across the infrastructure. Services are being deployed, services are having outages, databases are being resharded. With all of this change occurring, it can be difficult for a team to pinpoint the cause of a service outage. Digging through logs and dashboards is often insufficient.
Joey Parsons is the founder of Effx, a company that is building a platform for observing and managing the changes across the infrastructure. Effx is like a newsfeed for a service. An application instrumented with Effx gives the engineers a single endpoint that they can navigate to for understanding the history of their service.
Joey joins the show to talk about his experience as an infrastructure engineer at Airbnb, and how that experience informs the work of his new company, Effx.
A gig economy application generates lots of notifications.
There is SMS, mobile phone updates, emails, and native application updates. If you order a ride from Uber, you might receive a text message and a push notification at the same time. If an app overloads the user with notifications, the user might end up annoyed and delete the app from their phone.
But perhaps all of these notifications are necessary. You would rather get three simultaneous notifications from your food delivery app than fail to get your food on time. If you are the mobile application developer building the food delivery app, what other choice do you have?
At large companies such as Linkedin, there are entire teams devoted to figuring out how to optimize the notifications that they send you. It has a surprisingly large impact on the usability of a mobile application. Troy Goode is the founder of Courier, a company that provides notification optimization.
This might sound like a small, trivial problem. But it actually has a large impact on the usage of apps. And it is not an easy engineering problem. Troy joins the show to talk about the problem that Courier solves and the backend infrastructure that powers it. Courier is built entirely on serverless APIs. This is a great case study in how to build a completely scalable infrastructure product based on serverless tools.
In a modern data platform, distributed streaming systems are used to read data coming off of an application in real-time. There are a wide variety of streaming systems, including Kafka Streams, Apache Samza, Apache Flink, Spark Streaming, and more.
When Eric Anderson joined the show back in 2016, he was working at Google on Google Cloud Dataflow, a managed service for handling streaming data. Today, he works as an investor at Scale Venture Partners. In his current job, he analyzes companies built around data infrastructure, developer tooling, and other enterprise engineering domains.
Eric also hosts the podcast Contributor, which explores open source maintainers and the stories of their projects. His podcast has featured the creators of projects such as Envoy, Alluxio, and Chef. In today’s episode, Eric returns to the show to discuss data infrastructure, investing, and the evolving world of open source.
Distributed stream processing frameworks are used to rapidly ingest and aggregate large volumes of incoming data. These frameworks often require the application developer to write imperative logic describing how that data should be processed.
For example, a high volume of clickstream data that is getting buffered to Kafka needs to have a stream processing system evaluate that data to prepare it for a data warehouse, Spark, or some other queryable environment. In practice, many developers simply want to have the high volume of data become queryable in the fewest number of steps possible.
Materialize is a streaming SQL materialized view engine that provides materialized views over streaming data. The materialized views are incrementally updated over time and reconciled with new data that may have come in out of order.
Arjun Narayan and Frank McSherry are the co-founders of Materialize, a company whose technology is based on the Naiad paper, which was written at Microsoft Research. Arjun and Frank join the show to talk about modern streaming systems and their strategy for taking an academic paper and productizing it.
A data pipeline is a series of steps that takes large data sets and creates usable results from them. At the beginning of a data pipeline, a data set might be pulled from a database, a distributed file system, or a Kafka topic. Throughout a data pipeline, different data sets are joined, filtered, and statistically analyzed.
At the end of a data pipeline, data might be put into a data warehouse or Apache Spark for ad-hoc analysis and data science. At this point, the end-user of the data set expects that data to be clean and accurate. But how do we have any guarantees about the correctness?
Abe Gong is the creator of Great Expectations, a system for data pipeline testing. In Great Expectations, the developer creates tests called “expectations”, which verify certain characteristics of the data set at different phases in a data pipeline. This helps ensure that the end result of a multi-stage data pipeline is correct.
Abe joins the show to discuss the architecture of a data pipeline and the use cases of Great Expectations.
A data warehouse provides low latency access to large volumes of data.
A data warehouse is a crucial piece of infrastructure for a large company, because it can be used to answer complex questions involving a large number of data points. But a data warehouse usually cannot hold all of a company’s data at any given time. Users need to move a subset of the data into the data warehouse by reading large files from a data lake on disk and putting that data into the data warehouse.
The process of moving data from one place into another is broken down into three sequential steps, often called “ETL” (extract, transform, load) or “ELT” (extract, load, transform). In ETL, the data is extracted from a source such as a data lake, transformed into a schema that is customized for the data warehouse application, and then loaded into the data warehouse. In ELT, the last two steps are reversed, because modern systems can often leave the necessary schema transformation until after the data has been loaded into the data warehouse.
Matthew Scullion is the CEO of Matillion, a company that specializes in building tools for data transformations. Matthew joins the show to talk about the problem of data transformation, and how that problem has evolved over the nine years since he started Matillion.
If you enjoy the show, you can find all of our past episodes about data infrastructure by going to SoftwareDaily.com and searching for the technologies or companies mentioned. And if there is a subject that you want to hear covered, feel free to leave a comment on the episode, or send us a tweet @software_daily.
A data warehouse provides low latency access to large volumes of data.
A data warehouse is a crucial piece of infrastructure for a large company, because it can be used to answer complex questions involving a large number of data points. But a data warehouse usually cannot hold all of a company’s data at any given time. Users need to move a subset of the data into the data warehouse by reading large files from a data lake on disk and putting that data into the data warehouse.
The process of moving data from one place into another is broken down into three sequential steps, often called “ETL” (extract, transform, load) or “ELT” (extract, load, transform). In ETL, the data is extracted from a source such as a data lake, transformed into a schema that is customized for the data warehouse application, and then loaded into the data warehouse. In ELT, the last two steps are reversed, because modern systems can often leave the necessary schema transformation until after the data has been loaded into the data warehouse.
Matthew Scullion is the CEO of Matillion, a company that specializes in building tools for data transformations. Matthew joins the show to talk about the problem of data transformation, and how that problem has evolved over the nine years since he started Matillion.
If you enjoy the show, you can find all of our past episodes about data infrastructure by going to SoftwareDaily.com and searching for the technologies or companies mentioned. And if there is a subject that you want to hear covered, feel free to leave a comment on the episode, or send us a tweet @software_daily.
Machine learning applications are widely deployed across the software industry.
Most of these applications used supervised learning, a process in which labeled data sets are used to find correlations between the labels and the trends in that underlying data. But supervised learning is only one application of machine learning. Another broad set of machine learning methods is described by the term “reinforcement learning.”
Reinforcement learning involves an agent interacting with its environment. As the model interacts with the environment, it learns to make better decisions over time based on a reward function. Newer AI applications will need to operate in increasingly dynamic environments, and react to changes in those environments, which makes reinforcement learning a useful technique.
Reinforcement learning has several attributes that make it a distinctly different engineering problem than supervised learning. Reinforcement learning relies on simulation and distributed training to rapidly examine how different model parameters could affect the performance of a model in different scenarios.
Ray is an open source project for distributed applications. Although Ray was designed with reinforcement learning in mind, the potential use cases go beyond machine learning, and could be as influential and broadly applicable as distributed systems projects like Apache Spark or Kubernetes. Ray is a project from the Berkeley RISE Lab, the same place that gave rise to Spark, Mesos, and Alluxio.
The RISE Lab is led by Ion Stoica, a professor of computer science at Berkeley. He is also the co-founder of Anyscale, a company started to commercialize Ray by offering tools and services for enterprises looking to adopt Ray. Ion Stoica returns to the show to discuss reinforcement learning, distributed computing, and the Ray project.
If you enjoy the show, you can find all of our past episodes about machine learning, data, and the RISE Lab by going to SoftwareDaily.com and searching for the technologies or companies you are curious about . And if there is a subject that you want to hear covered, feel free to leave a comment on the episode, or send us a tweet @software_daily.
Machine learning applications are widely deployed across the software industry.
Most of these applications used supervised learning, a process in which labeled data sets are used to find correlations between the labels and the trends in that underlying data. But supervised learning is only one application of machine learning. Another broad set of machine learning methods is described by the term “reinforcement learning.”
Reinforcement learning involves an agent interacting with its environment. As the model interacts with the environment, it learns to make better decisions over time based on a reward function. Newer AI applications will need to operate in increasingly dynamic environments, and react to changes in those environments, which makes reinforcement learning a useful technique.
Reinforcement learning has several attributes that make it a distinctly different engineering problem than supervised learning. Reinforcement learning relies on simulation and distributed training to rapidly examine how different model parameters could affect the performance of a model in different scenarios.
Ray is an open source project for distributed applications. Although Ray was designed with reinforcement learning in mind, the potential use cases go beyond machine learning, and could be as influential and broadly applicable as distributed systems projects like Apache Spark or Kubernetes. Ray is a project from the Berkeley RISE Lab, the same place that gave rise to Spark, Mesos, and Alluxio.
The RISE Lab is led by Ion Stoica, a professor of computer science at Berkeley. He is also the co-founder of Anyscale, a company started to commercialize Ray by offering tools and services for enterprises looking to adopt Ray. Ion Stoica returns to the show to discuss reinforcement learning, distributed computing, and the Ray project.
If you enjoy the show, you can find all of our past episodes about machine learning, data, and the RISE Lab by going to SoftwareDaily.com and searching for the technologies or companies you are curious about . And if there is a subject that you want to hear covered, feel free to leave a comment on the episode, or send us a tweet @software_daily.
Distributed stream processing systems are used to read large volumes of data and perform operations across those data streams.
These stream processing systems often build off of the MapReduce algorithm for collecting and aggregating large volumes of data, but instead of processing a calculation over a single large batch of data, they process data on an ongoing basis. There are so many different stream processing system for this same use case–Storm, Spark, Flink, Heron, and many others.
Why is that? When there seems to be much more consolidation around the Hadoop MapReduce batch processing technology, why are there so many stream processing systems?
One explanation is that aggregating the results of a continuous stream of data is a process that very much depends on time. At any given point in time, you can take a snapshot of the stream of data, and any calculation based on that data is going to be out of date by the time that your calculation is finished. There is a latency between when you start calculating something, and when you finish calculating it.
There are other design decisions for a distributed stream processing system. What data do you keep in memory? What do you keep on disk? How often do you snapshot your data to disk? What is the method for fault tolerance? What are the APIs for consuming and processing this data?
Maximilian Michels has worked on the Apache Flink and Apache BEAM stream processing systems, and currently works on data infrastructure at Lyft. Max joins the show to discuss the tradeoffs of different stream processing systems and his experiences in the world of data processing.
You can find all of our past episodes about data infrastructure by going to SoftwareDaily.com and searching for the technologies or companies mentioned. And if there is a subject that you want to hear covered, feel free to leave a comment on the episode, or send us a tweet @software_daily.
From the publisher's feed