Improve Watch Queries

[stevensJourney]

Overview

This PR improves the watched query logic by optimizing table update handling, enhancing thread safety, and simplifying the implementation.

Table Update Optimization

• Table changes in a transaction are now tracked using a Set, preventing redundant entries. Since SQLite emits an update per row, the previous implementation could accumulate large arrays of duplicate table changes.

Thread Safety for Table Updates

• Resolves issue #135, where the current driver’s pending update storage was not thread-safe.
• The pending table changes set is now protected with Mutex locks.

Implementation Cleanup

• Replaces the existing SQLDelight watched query interface with a simpler, Flow-based solution.
• The new approach leverages the internal database’s getAll method instead of the driver’s synchronous methods.
• Table change flow logic has been refactored for better debouncing, allowing custom throttleMs values per query.

Removal of Blocking Calls

• Previously, the watch query API synchronously determined the tables to watch, which impacted performance (e.g., in the Supabase Todo List app). This logic now uses the asynchronous getAll method, improving efficiency.

Testing

This was tested locally using both the Kotlin and Swift SDKs.

watches.mp4

Swift

Related Swift PR powersync-ja/powersync-swift#26

[simolus3]

Thanks! I'm happy with the refactoring, and most of these comments are nits.

[simolus3]

LGTM!