cuVS-Lucene needs to be able to build a CAGRA graph out of core (larger than the amount of available GPU memory) without throwing OOM exceptions.
Elasticsearch has implemented a great strategy for enabling multiple concurrent index builds, while still being conscious of the memory usage. We should start by using that approach.
cuVS-Lucene needs to be able to build a CAGRA graph out of core (larger than the amount of available GPU memory) without throwing OOM exceptions.
Elasticsearch has implemented a great strategy for enabling multiple concurrent index builds, while still being conscious of the memory usage. We should start by using that approach.