Ytukyg

We have a microservice for relationship modeling between objects. A relation is defined between primary and secondary objects with cardinality constraints like 1-1, 1-N, N-N, etc.
The microservice provides API like Create relation, Find relations, Get secondaries, Get primaries, etc.

The query API "Get secondaries" takes a primary object and returns back all the related secondary objects. Since the related secondary object could be large, the results are paginated.

We had another microservice which was making good use of this relation microservice to work with relations. This consuming service accepted a similar pagination options like page index and number and passed the same to the relation service, and returned back the calling application the page results as obtained from the relation service. It was so far so good.

We recently identified that the consuming microservice was a bit chatty with the relation microservice as it had to call "Get secondaries" API multiple times given that there were multiple primary objects for which secondary objects had to be fetched.

So we thought to make the "Get Secondaries" API a bulk API by making it accept multiple primary objects as input. But then we got stuck with how the pagination would work.
The API would return related secondary objects for each primary but limit the secondary objects to the page size like earlier.

This seemed fine for the first call, but we are unsure how would this behave for the subsequent calls. If there were lesser number of secondary objects than the page size for one or more primary objects, what should be the input for subsequent calls. Do I need to pass those primary objects again?

This is where we are looking for suggestion on how to design this bulk API. Any input is welcome.

Basically, you should have some way to ensure that the relationship service knows what the original query was when receiving a paginated request.

A simple and maintainable way for your relationship service to handle this is to preprocess the request by sorting the requested primary objects in some way (ie. sort alphabetically by Id), and then simply iterating through the primary objects, adding secondary objects to the response, until the response is full.

The simplest thing for clients to do is to always use the same batch request and just add an index number or page token to the request.

I'd recommend a page token that mentions the last seen item (for example, lastSeen=primaryId,secondaryId (which you should obfuscate in some way to avoid a leaky abstraction)). Then, the service can look at the original request, and know where to resume iterating through all of the primary objects.

Alternately, you can encode enough information into a page token so that you can reconstruct whatever you need from the original request. This allows you to make some adjustments to the query on subsequent requests. (For example, if the client requests primaries A-Z, and you return secondary objects A1 - J5 in the first response, then you could modify the request to be J-Z; already seen J5, encode it so that you aren't leaking your implementation details, and return it to the client as the page token.) Then, instead of responding with the original request + page number, the client simply responds with the page token.

Either way, clients of the relationship service should never have to "figure out" what the request for the next page should be. The pagination should only require the consumer to increment a number or respond with a page token that was given to it by the relationship service.

Another consideration is the database that you are using. For example, in DynamoDB, the way to get the 100th item for a query like select * from secondaries where primaryId='ABC' requires you to read all items up to the 100th item. If you have a NoSQL database, or if you think you might move to a NoSQL database at some point in the future, you might find that a page token makes it much simpler to maintain where you are in the result set (as compared to an index number).

I found this article to be very helpful when I was learning about pagination myself, and I'd recommend reading it. It primarily deals with pagination concerns for UIs, but the fundamentals are the same.

TLDR: Don't make the consumer do any work. The consumer should repeat the original request with an added index number or page token, or the consumer should send a request containing only a page token.