Bloom filters, and when a probabilistic set is the right call
Trade a little certainty for a lot of memory — the math is friendlier than it looks.
Systems
KV cache, and why LLM inference is memory-bound
The cache that makes autoregressive decoding fast also makes it the thing that runs out of memory first.
Notes on cgroups v2: memory limits that actually hold
Why my OOM kills moved around after switching to the unified hierarchy, and the three knobs that matter.