Migrating from Hotvect v7 to Hotvect v8¶

This guide provides information on migrating from Hotvect v7 to Hotvect v8.

Why You Should Migrate¶

Improved TopK Support¶

Algorithm developers can now attach pre-computed ML features to the AvailableActions to be used in a TopK request. I.e. the AvailableActionState can now serve as a feature store (for the candidates to be considered in TopK operations). For example, vector representations of the available actions can be stored, which can then be used for vector search etc.

Pre-computed features are seamlessly integrated into the memoization framework, allowing them to be accessed in the same way as other features.

Enhanced API¶

The memoization API introduced in v7 has been generalized into a computation API, which covers computation without caching, memoization (caching within the scope of a request), and pre-computation (caching across requests). As a result, the memoization API has been moved from the hotvect-core module to the hotvect-api module. At the same time, mMethods and classes have been renamed to be more intuitive.

Support for Multiple Data Sources in State Generation¶

The GenerateState interface now supports multiple data sources, providing greater flexibility when generating state.

Improved State Codec Interface¶

Previously, the StateCodec interface required the same type for both deserialization and serialization. This restriction has been removed; algorithm developers can now use different types for serialization and deserialization. In future versions, we may require developers to implement only the serialization method, while deserialization is left to the algorithm developer (i.e. outside of the framework).

Better Module Isolation¶

The initial design intended for the hotvect-core module to be specific to algorithms, allowing different algorithms to use different versions of core modules. This isolation was compromised in practice, as the core module was included in the hotvect-offline module, creating potential version discrepancies between online and offline environments. In v8, this has been fixed. The hotvect-core module is now excluded from both the hotvect-online and hotvect-offline modules (such that it is only provided by the algorithm jar).

`PredictStdOut` Command-Line Tool¶

A new command-line tool, PredictStdOut, has been added to the hotvect-offline module. This tool allows for predictions to be performed in an interactive manner, aiding in debugging processes.

File Aggregator with Enhanced Performance¶

The “Unordered” file aggregator has been introduced. Unordered processing generally offers better performance than ordered processing, though compared to the “Unordered” file mapper, it may be more susceptible to bottlenecks due to the operations on the shared aggregated state.

Bug Fixes and Other Improvements¶

Removal of the `offline-online-analysis` Module¶

This module was poorly designed and unused in practice; it has been removed.

Dropping Support for v6¶

Support for the v6 API has been removed.

Known Issues¶

Due to the removal of the hotvect-core module from the hotvect-offline module, the commandline task for listing available transformations is currently broken (removed). This issue will be addressed in future versions.