Alternatives to Apache Spark

The early data ingestion pipeline at Pinterest used Kafka as the central message transporter, with the app servers writing messages directly to Kafka, which then uploaded log files to S3.

For databases, a custom Hadoop streamer pulled database data and wrote it to S3.

Challenges cited for this infrastructure included high operational overhead, as well as potential data loss occurring when Kafka broker outages led to an overflow of in-memory message buffering.

Why we built Marmaray, an open source generic data ingestion and dispersal framework and library for Apache Hadoop :

Built and designed by our Hadoop Platform team, Marmaray is a plug-in-based framework built on top of the Hadoop ecosystem. Users can add support to ingest data from any source and disperse to any sink leveraging the use of Apache Spark . The name, Marmaray, comes from a tunnel in Turkey connecting Europe and Asia. Similarly, we envisioned Marmaray within Uber as a pipeline connecting data from any source to any sink depending on customer preference:

https://eng.uber.com/marmaray-hadoop-ingestion-open-source/

(Direct GitHub repo: https://github.com/uber/marmaray Kafka Kafka Manager )

I am designing a Django application for my organization which will be used as an internal tool. The infra team said that I will not be having SSH access to the production server and I will have to log all my backend application messages to Splunk. I have no knowledge of Splunk so the following are the approaches I am considering: Approach 1: Create an hourly cron job that uploads the server log file to some Splunk storage for later analysis. - Is this possible? Approach 2: Is it possible just to stream the logs to some splunk endpoint? (If yes, I feel network usage and communication overhead will be a pain-point for my application)

Is there any better or standard approach? Thanks in advance.

I use Kibana because it ships with the ELK stack. I don't find it as powerful as Splunk however it is light years above grepping through log files. We previously used Grafana but found it to be annoying to maintain a separate tool outside of the ELK stack. We were able to get everything we needed from Kibana.

After years of optimizing our existing feed technology, we decided to make a larger leap with 2.0 of Stream. While the first iteration of Stream was powered by Python and Cassandra, for Stream 2.0 of our infrastructure we switched to Go.

The main reason why we switched from Python to Go is performance. Certain features of Stream such as aggregation, ranking and serialization were very difficult to speed up using Python.

We’ve been using Go since March 2017 and it’s been a great experience so far. Go has greatly increased the productivity of our development team. Not only has it improved the speed at which we develop, it’s also 30x faster for many components of Stream. Initially we struggled a bit with package management for Go. However, using Dep together with the VG package contributed to creating a great workflow.

Go as a language is heavily focused on performance. The built-in PPROF tool is amazing for finding performance issues. Uber’s Go-Torch library is great for visualizing data from PPROF and will be bundled in PPROF in Go 1.10.

The performance of Go greatly influenced our architecture in a positive way. With Python we often found ourselves delegating logic to the database layer purely for performance reasons. The high performance of Go gave us more flexibility in terms of architecture. This led to a huge simplification of our infrastructure and a dramatic improvement of latency. For instance, we saw a 10 to 1 reduction in web-server count thanks to the lower memory and CPU usage for the same number of requests.

#DataStores #Databases

1.0 of Stream leveraged Cassandra for storing the feed. Cassandra is a common choice for building feeds. Instagram, for instance started, out with Redis but eventually switched to Cassandra to handle their rapid usage growth. Cassandra can handle write heavy workloads very efficiently.

Cassandra is a great tool that allows you to scale write capacity simply by adding more nodes, though it is also very complex. This complexity made it hard to diagnose performance fluctuations. Even though we had years of experience with running Cassandra, it still felt like a bit of a black box. When building Stream 2.0 we decided to go for a different approach and build Keevo. Keevo is our in-house key-value store built upon RocksDB, gRPC and Raft.

RocksDB is a highly performant embeddable database library developed and maintained by Facebook’s data engineering team. RocksDB started as a fork of Google’s LevelDB that introduced several performance improvements for SSD. Nowadays RocksDB is a project on its own and is under active development. It is written in C++ and it’s fast. Have a look at how this benchmark handles 7 million QPS. In terms of technology it’s much more simple than Cassandra.

This translates into reduced maintenance overhead, improved performance and, most importantly, more consistent performance. It’s interesting to note that LinkedIn also uses RocksDB for their feed.

#InMemoryDatabases #DataStores #Databases

I have to build a data processing application with an Apache Beam stack and Apache Flink runner on an Amazon EMR cluster. I saw some instability with the process and EMR clusters that keep going down. Here, the Apache Beam application gets inputs from Kafka and sends the accumulative data streams to another Kafka topic. Any advice on how to make the process more stable?

Lumosity is home to the world's largest cognitive training database, a responsibility we take seriously. For most of the company's history, our analysis of user behavior and training data has been powered by an event stream--first a simple Node.js pub/sub app, then a heavyweight Ruby app with stronger durability. Both supported decent throughput and latency, but they lacked some major features supported by existing open-source alternatives: replaying existing messages (also lacking in most message queue-based solutions), scaling out many different readers for the same stream, the ability to leverage existing solutions for reading and writing, and possibly most importantly: the ability to hire someone externally who already had expertise.

We ultimately migrated to Kafka in early- to mid-2016, citing both industry trends in companies we'd talked to with similar durability and throughput needs, the extremely strong documentation and community. We pored over Kyle Kingsbury's Jepsen post (https://aphyr.com/posts/293-jepsen-Kafka), as well as Jay Kreps' follow-up (http://blog.empathybox.com/post/62279088548/a-few-notes-on-kafka-and-jepsen), talked at length with Confluent folks and community members, and still wound up running parallel systems for quite a long time, but ultimately, we've been very, very happy. Understanding the internals and proper levers takes some commitment, but it's taken very little maintenance once configured. Since then, the Confluent Platform community has grown and grown; we've gone from doing most development using custom Scala consumers and producers to being 60/40 Kafka Streams/Connects.

We originally looked into Storm / Heron , and we'd moved on from Redis pub/sub. Heron looks great, but we already had a programming model across services that was more akin to consuming a message consumers than required a topology of bolts, etc. Heron also had just come out while we were starting to migrate things, and the community momentum and direction of Kafka felt more substantial than the older Storm. If we were to start the process over again today, we might check out Pulsar , although the ecosystem is much younger.

To find out more, read our 2017 engineering blog post about the migration!

When I joined NYT there was already broad dissatisfaction with the LAMP (Linux Apache HTTP Server MySQL PHP) Stack and the front end framework, in particular. So, I wasn't passing judgment on it. I mean, LAMP's fine, you can do good work in LAMP. It's a little dated at this point, but it's not ... I didn't want to rip it out for its own sake, but everyone else was like, "We don't like this, it's really inflexible." And I remember from being outside the company when that was called MIT FIVE when it had launched. And been observing it from the outside, and I was like, you guys took so long to do that and you did it so carefully, and yet you're not happy with your decisions. Why is that? That was more the impetus. If we're going to do this again, how are we going to do it in a way that we're gonna get a better result?

So we're moving quickly away from LAMP, I would say. So, right now, the new front end is React based and using Apollo. And we've been in a long, protracted, gradual rollout of the core experiences.

React is now talking to GraphQL as a primary API. There's a Node.js back end, to the front end, which is mainly for server-side rendering, as well.

Behind there, the main repository for the GraphQL server is a big table repository, that we call Bodega because it's a convenience store. And that reads off of a Kafka pipeline.

To provide employees with the critical need of interactive querying, we’ve worked with Presto, an open-source distributed SQL query engine, over the years. Operating Presto at Pinterest’s scale has involved resolving quite a few challenges like, supporting deeply nested and huge thrift schemas, slow/ bad worker detection and remediation, auto-scaling cluster, graceful cluster shutdown and impersonation support for ldap authenticator.

Our infrastructure is built on top of Amazon EC2 and we leverage Amazon S3 for storing our data. This separates compute and storage layers, and allows multiple compute clusters to share the S3 data.

We have hundreds of petabytes of data and tens of thousands of Apache Hive tables. Our Presto clusters are comprised of a fleet of 450 r4.8xl EC2 instances. Presto clusters together have over 100 TBs of memory and 14K vcpu cores. Within Pinterest, we have close to more than 1,000 monthly active users (out of total 1,600+ Pinterest employees) using Presto, who run about 400K queries on these clusters per month.

Each query submitted to Presto cluster is logged to a Kafka topic via Singer. Singer is a logging agent built at Pinterest and we talked about it in a previous post. Each query is logged when it is submitted and when it finishes. When a Presto cluster crashes, we will have query submitted events without corresponding query finished events. These events enable us to capture the effect of cluster crashes over time.

Each Presto cluster at Pinterest has workers on a mix of dedicated AWS EC2 instances and Kubernetes pods. Kubernetes platform provides us with the capability to add and remove workers from a Presto cluster very quickly. The best-case latency on bringing up a new worker on Kubernetes is less than a minute. However, when the Kubernetes cluster itself is out of resources and needs to scale up, it can take up to ten minutes. Some other advantages of deploying on Kubernetes platform is that our Presto deployment becomes agnostic of cloud vendor, instance types, OS, etc.

#BigData #AWS #DataScience #DataEngineering

I have to collect different data from multiple sources and store them in a single cloud location. Then perform cleaning and transforming using PySpark, and push the end results to other applications like reporting tools, etc. What would be the best solution? I can only think of Azure Data Factory + Databricks. Are there any alternatives to #AWS services + Databricks?