One of the most significant parts of the Dataloop platform is the data operations engine, which powers their pipeline creation, automation and Function-as-a-Service features. This is what allows their customers to blend code, data, etc. into one cohesive, harmonic pipeline. Using this pipeline, customers can then insert triggers or filters that split information into tasks — annotate, validate, train a model with, pass into a different dataset or create an automation.

This level of data processing requires a great deal of cloud computing power, but forecasting that usage proved to be an impossible task given Dataloop customers’ unpredictable needs. One day, a customer might be completely silent as they build their pipeline, but the next they might upload over a million items or have 1,000 annotators on the platform. These fluctuations make it challenging for Dataloop to determine how much compute power they’ll need in the following week, let alone in future months.

Given that uncertainty, purchasing a 1-year or multiyear CUD from Google Cloud becomes very risky, so Dataloop purchased one with a conservative commitment threshold. However, internal cloud optimization efforts lowered their compute needs to about 70% of their CUD commitment, which in turn prompted underutilization fees. Essentially, they were being penalized for becoming more efficient.

Dataloop needed a way to manage their on-demand instances that would accommodate their customers’ usage bursts but would keep cost effectiveness.. An ideal solution would be easy and quick to implement, thus freeing time to work on actually expanding and growing the company’s service offerings.