once a server pool is initialized, its storage layout, based on consistent hashing, becomes immutable. this design eliminates the need to consider dynamic node additions or removals within a cluster, thereby completely avoiding the complexities of rebalancing, allowing a no-metadata design. 🤔

we can see a similar philosophy at play with MinIO's server pools. it's somewhat counter-intuitive that you can't simply expand the capacity of an existing MinIO cluster (server pool). instead, to scale, you must add an entirely new, fixed-size cluster (server pool) to the system.

in an LWS cluster, the leader node first start up to awaiting the joins of the worker nodes. if even a single node fails, the entire cluster is rebuilt. my understanding is that this design simplifies dynamic, continuous membership management into a one-time operation at the initliazation phase. 🤔

for instance, LLM inference systems like vLLM often leverage collective communication (e.g. NCCL), and once a cluster is established, its topology remains unchanged. in Kubernetes, a LWS (Leader Worker Set) is particularly well-suited for this type of workload. 🤔

when scaling is required, one simply continues to add fixed-size clusters to the system, and put some load balacing in front of multiple clusters.

the truth has been spoken, the people who can fund the housework robots don't have to do the dishes themselves 😂

if robots could handle these logistical tasks, a significant portion of housework would already be freed up. the form of the robot, whether humanoid or not, probably isn't that important. 🤔

thanks! hope it can be useful 😁

yeah, it's an optimistic numbering scheme, if there's any username conflict, I'll just back off to the next year and try to register the account again. 😂

He will give practical advice, and concrete criteria to consider, when choosing research projects, and making professional decisions, in these last few years before AGI."

docs.google.com/presentation...