A study quantified how large context windows degrade LLM performance on coding tasks using GPT-5.6 Sol and Claude Opus 5. GPT-5.6 Sol showed significant performance drops at 250k+ tokens regardless of context relevance, while Claude Opus 5 only degraded when context was task-related. The models use different tokenizers, requiring normalization to UTF-8 MB for fair comparison.
A Kafka cluster experienced severe performance degradation when eight single-member consumer groups, each fetching 4,742 partitions, caused lock contention in the broker's DelayedFetch mechanism. The issue emerged after upgrading to Kafka 2.7+, which changed lock acquisition to blocking mode, starving request handlers and blocking network threads while broker-measured latency metrics remained artificially low because time spent in unread sockets goes unmeasured.