Issue Grouping: Smarter, Faster, Half as Wrong
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
Sentry upgraded its AI-powered issue grouping model from v1 to v2, achieving a 20% increase in duplicate issue prevention and cutting incorrect merges (overgrouping) from 8% to 4%. The new model uses ModernBERT trained on hundreds of thousands of Claude-labeled stacktrace pairs, replacing Jina AI's off-the-shelf embeddings. Key inference improvements include Matryoshka Representation Learning with embedding truncation from 768 to 64 dimensions (4x faster HNSW lookups, 12x less DB storage), bfloat16 precision, PyTorch SDPA, and CUDA graph compilation for ~3x lower inference latency. The migration strategy avoided a full backfill by using a live fallback chain between v1 and v2 embeddings, with a training_mode flag to backfill v2 representations from live traffic. The full rollout took 6 weeks with zero customer-reported regressions.
