OpenAI Coding Interview: Time-Based GPU Credit System

Implement a replayable, time-based credit tracking service that models GPU credits valid on half-open intervals [start, expiration). You must support three operations: record credit grants (adds), attempt charges (consumes credits at an event timestamp), and query the balance at an arbitrary timestamp. All operations are driven by event timestamps and the system must tolerate out-of-order arrival: queries consider only events with timestamp ≤ query time.

During a balance query you must replay cached events from the first recorded event up to the query time, applying adds as grants valid on [start, expiration) and applying charges by consuming from grants that are added ≤ charge time and unexpired at that time. When consuming, prefer grants that expire earliest. Charges that would make balance negative must fail (return None) and not modify state for events ≤ that timestamp. If replay finds recorded charges that cannot be reconciled with available grants, get_balance must raise an exception.

Flow you should expect in an interview: specify method signatures and side effects, describe data structures for cached events and grant bookkeeping, implement correct replay semantics, add persistence/restore, then discuss optimizations and trade-offs.

Skill signals: demonstrate event-sourcing thinking, sweep-line / interval handling, greedy consumption by earliest-expiry, correctness under out-of-order events, deterministic persistence format, and trade-offs for scaling (snapshotting, indexing, incremental aggregates). Boldly document any public structures and the persisted byte format so the harness can restore identical semantics.