This article discusses a solution to inconsistent search result rankings caused by distributed search replicas. By replacing the standard BM25 scoring function with a new implementation called StableTfl, the authors eliminated ranking drift without requiring additional state or index snapshots. StableTfl ensures consistent results across replicas by relying on static constants rather than corpus-dependent statistics.
Key points
Distributed search replicas often produce inconsistent ranking results because standard algorithms like BM25 rely on corpus statistics that vary between nodes.
Stateful workarounds for pagination consistency, such as session pinning, introduce significant architectural complexity.
StableTfl provides deterministic scores by calculating relevance based solely on the query and document, removing dependency on distributed corpus metrics.
The trade-off for deterministic consistency is a slight reduction in absolute retrieval quality compared to BM25, necessitating a context-dependent choice between the two.