Takeaways from https://engineering.linkedin.com/blog/2017/08/scaling-contextual-conversation-suggestions-over-linkedins-graph
Data jiujitsu: When building a data product, start small and verify that the users like the idea before investing too much to get a perfect product. This is reiterated in “Rules of Machine Learning: Best Practices for ML Engineering” by Martin Zinkevich.
Hadoop joins: If your offline flow is taking a long time to converge, it might be that you are doing massive joins.
Use hybrid: When building an online recommendation service, consider using a hybrid solution to precompute some parts of the computation in order to speed up your service. We have other successful systems at LinkedIn that follow a similar approach, including our Ads ML system.