Friend feature

[avinh]

How to ?

• Get a list of all the user's friends
• Check if 2 people are friends
  Can this model handle 9 million users? how is the performance

[kyleishie]

Answering your questions in order.

1. I have yet to build a traversal API for this. My initial and admittedly naive intention is that traversal will allow at least one hop by simply querying the entire partition of a known vertex. If you look at how edges are stored in DDB you will notice that creating an edge creates a copy of v2 into the v1 partition. This copy is intended to be read-only and to be kept in sync with the original v2... eventually. 😁 This read-only copy of v2 provides an in memory version for the edge to point to and removes the need to perform a DDB query for one hop. The edge is also written into the v2 partition as a reciprocal. My intention is to use that later for traversal to support out vs in hops. So to answer your question... Right now you can't because the API doesn't exist, however, you could build the DDB query yourself for now.

2. Checking for the existence of an edge between two vertices could be built as a DDB query within the v1 partition or v2 partition depending on the expected direction of the edge. For example, in the context of person1, person2, and a friendship edge, one could perform the following DDB query. While this does NOT return the person2 object itself it does confirm the existence of the edge.
   { __p : "PERSON#person1", __s : ">friendship#PERSON#person2" }

3. Since there is no traversal API yet, the API in its current form is basically just DDB getItem and putItem calls, therefore, there is no meaningful performance metric to test.

Please keep in mind that this library is a proof of concept I put together to get a half-baked idea out on paper. This is by no means a production ready library. I will continue to work on it until the idea is proven wrong but I do not recommend using it for anything other than trying to prove it wrong or improve it.

[avinh]

I make a contribution like this, you create too many copies item so a table containing too much data leads to slow queries and heavy data, which will make maintenance costs high. I have just learned about Dynamodb and would like to have a cost-effective solution

Hope you will continue to develop it

[kyleishie]

I completely understand that this is very bandwidth and storage heavy. The idea (or better the challenge) was to come up with a primary key schema that would allow at least one hop per query. One thing to note here is that due to the nature of DynamoDB primary key it is impossible to have more than one copy of a v2 within a v1 partition. The solution I have here leaves a lot to be desired and that is the major reason I have yet to implemented traversal.

Here is what I was thinking:

1. DynamoDB is the persistent storage mechanism for the graph.
2. A worker (lambda, container, etc.) could query an entire partition into memory when needed. The initial query is slow (similar to a lambda cold start) but all subsequent queries into the same partition would be in memory.
3. The in memory representation is ephemeral (for example, kept in memory for the life time of a lambda execution) and therefore eventually consistent in the most literal sense.

Again there is a lot to be desired here... 😅

I may make this a part time project soon. I am current in the process of moving to a new home and do not have much are time outside of work. I would be happy to discuss any ideas you have about this here. 😁