Grafana Mimir vs parglm

parglm fits generalized linear models using parallel QR decomposition, targeting datasets where glm() is too slow. Tom Palmer took over maintenance at 0.1.8 in April 2026, and the package has released five times since — a burst of activity after a long quiet period. 0.2.0 in July 2026 is the first release to focus on correctness rather than throughput.

The arc runs from performance to trustworthiness. 0.1.9 was a large optimization release — deque-based task queues, fused memory passes, upper-triangle-only Fisher information, thread_local IDs — plus ecosystem integration with sandwich and gtsummary. 0.2.0 then fixed an out-of-bounds write triggered by small block_size values and a path where a non-finite working response could poison the QR decomposition, and brought response-type handling in line with glm().

With the memory-safety issues addressed and glm parity closed for binomial responses, further work is likely to extend family coverage or the benchmark suite rather than revisit the threading model. The C++17 requirement set at 0.1.8 gives room for more aggressive optimization if the maintainer returns to that.