Alternatives to NServiceBus

As Sentry runs throughout the day, there are about 50 different offline tasks that we execute—anything from “process this event, pretty please” to “send all of these cool people some emails.” There are some that we execute once a day and some that execute thousands per second.

Managing this variety requires a reliably high-throughput message-passing technology. We use Celery's RabbitMQ implementation, and we stumbled upon a great feature called Federation that allows us to partition our task queue across any number of RabbitMQ servers and gives us the confidence that, if any single server gets backlogged, others will pitch in and distribute some of the backlogged tasks to their consumers.

#MessageQueue

We've been using RabbitMQ as Zulip's queuing system since we needed a queuing system. What I like about it is that it scales really well and has good libraries for a wide range of platforms, including our own Python. So aside from getting it running, we've had to put basically 0 effort into making it scale for our needs.

However, there's several things that could be better about it: * It's error messages are absolutely terrible; if ever one of our users ends up getting an error with RabbitMQ (even for simple things like a misconfigured hostname), they always end up needing to get help from the Zulip team, because the errors logs are just inscrutable. As an open source project, we've handled this issue by really carefully scripting the installation to be a failure-proof configuration (in this case, setting the RabbitMQ hostname to 127.0.0.1, so that no user-controlled configuration can break it). But it was a real pain to get there and the process of determining we needed to do that caused a significant amount of pain to folks installing Zulip. * The pika library for Python takes a lot of time to startup a RabbitMQ connection; this means that Zulip server restarts are more disruptive than would be ideal. * It's annoying that you need to run the rabbitmqctl management commands as root.

But overall, I like that it has clean, clear semanstics and high scalability, and haven't been tempted to do the work to migrate to something like Redis (which has its own downsides).

The algorithms and data infrastructure at Stitch Fix is housed in #AWS. Data acquisition is split between events flowing through Kafka, and periodic snapshots of PostgreSQL DBs. We store data in an Amazon S3 based data warehouse. Apache Spark on Yarn is our tool of choice for data movement and #ETL. Because our storage layer (s3) is decoupled from our processing layer, we are able to scale our compute environment very elastically. We have several semi-permanent, autoscaling Yarn clusters running to serve our data processing needs. While the bulk of our compute infrastructure is dedicated to algorithmic processing, we also implemented Presto for adhoc queries and dashboards.

Beyond data movement and ETL, most #ML centric jobs (e.g. model training and execution) run in a similarly elastic environment as containers running Python and R code on Amazon EC2 Container Service clusters. The execution of batch jobs on top of ECS is managed by Flotilla, a service we built in house and open sourced (see https://github.com/stitchfix/flotilla-os).

At Stitch Fix, algorithmic integrations are pervasive across the business. We have dozens of data products actively integrated systems. That requires serving layer that is robust, agile, flexible, and allows for self-service. Models produced on Flotilla are packaged for deployment in production using Khan, another framework we've developed internally. Khan provides our data scientists the ability to quickly productionize those models they've developed with open source frameworks in Python 3 (e.g. PyTorch, sklearn), by automatically packaging them as Docker containers and deploying to Amazon ECS. This provides our data scientist a one-click method of getting from their algorithms to production. We then integrate those deployments into a service mesh, which allows us to A/B test various implementations in our product.

For more info:

• Our Algorithms Tour: https://algorithms-tour.stitchfix.com/
• Our blog: https://multithreaded.stitchfix.com/blog/
• Careers: https://multithreaded.stitchfix.com/careers/

#DataScience #DataStack #Data

As we've evolved or added additional infrastructure to our stack, we've biased towards managed services. Most new backing stores are Amazon RDS instances now. We do use self-managed PostgreSQL with TimescaleDB for time-series data—this is made HA with the use of Patroni and Consul.

We also use managed Amazon ElastiCache instances instead of spinning up Amazon EC2 instances to run Redis workloads, as well as shifting to Amazon Kinesis instead of Kafka.

To solve the problem of scheduling and executing arbitrary tasks in its distributed infrastructure, PagerDuty created an open-source tool called Scheduler. Scheduler is written in Scala and uses Cassandra for task persistence. It also adds Apache Kafka to handle task queuing and partitioning, with Akka to structure the library’s concurrency.

The service’s logic schedules a task by passing it to the Scheduler’s Scala API, which serializes the task metadata and enqueues it into Kafka. Scheduler then consumes the tasks, and posts them to Cassandra to prevent data loss.

I decided to use Akka instead of Kafka streams because I have personal relationships at @Lightbend.