Hi Dennis, that's a good question!
GSI is a viable solution, yes. It would simplify the implementation, but also comes with its shortcomings.
GSIs can't offer consistency. There will always be a delay between changes to the table and reflections in the secondary index. For some use cases, this is unbearable. With the approach I suggested, it's possible to wrap many item operations into one transaction, what gives you full control and consistency.
Another issue with GSI is that the developer now has two sources for the same dataset and needs to discover on which to read data from.
Consider the Friendly example. If you want to find all Ross' siblings, do you need to query the table, the GSI or both? If your data structure is really really simple, this might not be a problem. Otherwise, it will complicate your life when implementing new features that need to read data. And can easily end up in a mess. You could start missing data that exists by querying the wrong source, and showing incomplete information to your users. And you may never find out...
In order to minimize errors when writing data using the approach I outlined, you could have a single internal service for entering nodes and edges in the DB. You take care of this single entry point and make sure it's handing all relationships correctly. Then all other parts of your application will only invoke this single service to write in the DB.
Does that make sense?
Hmmm, not sure I'd use DDB at all if eventual consistency was a problem, as I he transaction operations throws a lot of performance away (and are somewhat limited), but fair enough.
As we're getting to the connections between equals, like the siblings example, you'd need somewhat arbitrary rules (e.g. eldest has the relation), or do the copying anyway, and then we'd both have the extra redundancy AND the GSI to worry about, so the triplestore model makes a lot of sense there.
I'm not sure that I'm convinced for data with a clearer parent-child relationship, I'll have to let it simmer for a while. :-)
By the way, you're storing the predicates in the DB, but doesn't really use them for anything. I guess that's just for discovery (e.g. when using it as storage for clients that doesn't necessarily know what relation types that exist), but are there other uses for those in a less generalised database? They seem a little bit useless to me unless you add an attribute to make a sparse index from (otherwise you'd have to do a full table scan to find them), but I've probably missed something.
Thanks for the clarification!
We're a place where coders share, stay up-to-date and grow their careers.
We strive for transparency and don't collect excess data.