Flávio Juvenal

As this is buggy in current main/master branch, as explained by @dr-ftvkun above, I also think this should be merged.

Any easy workaround instead of passing all excludes as a CLI argument? This issue broke several of my CIs.

Make sure it's available in frontend too.

@gabicavalcante One of our projects is using CONN_MAX_AGE = 600 w/o problem in production since last week. > We need to be careful because If the database terminates idle connections...

Thanks @fgregg, I'll investigate. With current implementation, I had memory errors at this line, when numpy needs to materialize part of the mmap-ed `scored_pairs/edgelist` array into main memory because of...

@fgregg Good to know! I can't share it because it's private data, but I'll try to generate one.

@fgregg so even after avoiding the random access at `sub_graph = edgelist[component]` (with `sub_graph = edgelist[start:stop]`) the bottleneck still exists? Are there any offending lines or the whole algorithm is...

but how efficient would be to grab edges from vertices? Seems to be a memory locality trade-off

graphframes library (Spark-based) uses [this algorithm](https://youtu.be/Io1x6mQlh1E). Related issue is here https://github.com/graphframes/graphframes/issues/117

AFAIK, the problem with multinomial classification is how to have flexibility between training and evaluation on the number N of classes. So instead of training with N fixed classes, deep...