Predict and train.
Chalk supports online and offline resolvers. Online resolvers produce features for online inference, and offline resolvers produce offline training data.
Online and offline resolvers have different read/write needs. Your online environment stores data that you want to cache for the period you’d like to cache it. While this is usually a small amount of data, you want exceptionally fast reads. These access patterns fit well with the performance characteristics of Redis or BigTable, where Chalk stores your online data.
The offline environment stores far more data than the online environment. It keeps a record of all online runs and indexes all data brought in from your offline data sources. This timeseries data is highly compressible, as many features change less frequently than they are sampled (for example, users on your platform might very rarely change their names.) Chalk uses the column store Timescale or the data warehouse BigQueryfor this data.
Online data is used to receive information about a single entity. For example, you might be looking to compute the features of a credit model for a single user, or decide what products to suggest to a customer. You can use our API client to pull this information.
Offline data is used to query historical data about one or more entities at specific points in time. Offline data is used to train models or do investigative work. See our guide on querying training data for a more in-depth treatment.
Both offline and online resolvers support scheduled runs. However, you’re likely to consider scheduling more frequently with offline resolvers. Every offline resolver pulls in data from your underlying source to Chalk’s feature store on a schedule. Chalk will determine the frequency at which to poll your data sources, or you can choose to provide a custom duration on which to pull the data.
Online resolvers can also be scheduled, although they don’t necessarily run on a schedule as offline resolvers do.
Both online and offline scheduled resolvers may take arguments, and you have control over the sets of arguments to run.
After an online resolver runs, its values are copied into the offline store. When you query the offline store, you will receive data from both records of online runs and offline-specific resolvers. Which data you receive depends on which data was closest to the point-in-time that you queried. For more information, see temporal consistency.
In contrast, data from the offline environment does not reach the online store by default. However, you can choose to ETL the data from an offline resolver into the online store. This can be helpful, for example, when you tolerate stale data in online inference and have a data source in the offline environment that doesn’t have a direct replacement in the online environment. More details are provided in the section Reverse ETL.