Petabyte-Scale Data Pipelines with Docker, Luigi and Elastic Spot Instances

tech.adroll.com

51 points·vtuulos··5 comments

Cool stuff. Curious to know what systems these containerized tasks are pulling data from. Does adroll consider let those containers access production database instances? Or are they backed by non-production systems? (EDIT: I understand that it's S3 for intermediate steps. But I'm curious where the data comes from initially.)

We have a few cases of containers pulling (meta)data from databases. We use Postgres on RDS, so setting up a read-replica is very easy. Containers can safely access these replicas.

Raw data is collected by our bidders and adservers that push data to S3 and Kinesis for real time consumers (http://tech.adroll.com/blog/data/2015/06/26/kinesis.html)

Interesting, we're doing similar things with Spark, Cassandra and Mesos. By scaling out Mesos with spot instances for its agents.

What are you using for scheduling your jobs?

We wrote our own scheduler, called Quentin, which we may opensource. Essentially it is a simple queue that takes care of running containers and managing pools of instances that containers run on by integrating with AWS autoscaling.

It is a very straightforward service, for example it doesn't do retries (done by Luigi), it is not highly available and not distributed. That makes things much easier to maintain, and we don't need HA since jobs inputs/outputs are in S3 anyway. Luigi makes it easy to restart parts of the pipeline if scheduler ever goes down.

Reply on news.ycombinator.com