I suggest the Search Platform team evaluate whether it is feasible to have 2 different backends instead of one.
The rationale is as follows:
- users could choose between 2 different WDQS endpoints running on different backends
- it gives WMF a possibility to measure and compare performance over time
- it makes it possible to later on discontinue the least suitable of the 2
- it makes it less likely to end up in a Blazegraph situation again
- we don't have a shortage of computing resources or money, so 2 clusters with 6 machines each is fine if needed
- users will themselves distribute between the 2 endpoints and help lower the load for each cluster
- it will be easier to conduct experiments on one cluster at a time as long as the other is up (redundancy)
Evaluate whether one of the services could be updated every 48-72 hours and have other benefits (like speed and better precomputed optimizations) than the "live" one that tracks edits in near-real time.
see https://ad-publications.cs.uni-freiburg.de/ARXIV_sparql_autocompletion_BKKKS_2021.pdf for a paper mentioning optimizations of e.g. wdt:P31/wdt:P279*-queries which are very valuable and common at least to me (but often time out on WDQS and most probably will be too expensive also on a future BG replacement).