Rate limiting and pagination aren’t (necessarily) about making full data consumption more difficult. They’re more often about optimizing common use cases and general quality of service.
Edit to add: in certain circles (eg those of us who take REST and HATEOAS as baseline HTTP API principles), parallelism is often not just expected but often encouraged. A service can provide efficient, limited subsets of a full representation and allow clients to retrieve as little or as much of the full representation as they see fit.
But limiting for efficiency is usually done in a way that I would call a cargo cult: First, the number of items per "page" is usually a number one would pick per displayed page, in the range of 10 to 20. This is inefficient for the general case, the amount of data transmitted is usually just the same size as the request plus response headers. So if the API isn't strictly for display purposes, pick a number of items per page that gives a useful balance between not transmitting too much useless data, but keeping query and response overhead low. Paginate in chunks of 100kB or more.
In terms of computation and backend load, pagination can be as expensive for a 1-page-query as for a full query. Usually this occurs when the query doesn't directly hit an index or similar data structure where a full sweep over all the data cannot be avoided. So think and benchmark before you paginate, and maybe add an index here and there.
Pagination would just complicate things. I think with most APIs, intended as APIs (i.e., not just an endpoint primarily meant to feed a front-end page), you're better off thinking of your default as "I'm going to just stream everything they ask for", and look for reasons why that won't work, rather than start from the presumption that everything must be paginated from the beginning.
Don't get me wrong; there are plenty of solid reasons to paginate. You may discover one applies to your API. But if you can leave it out, it's often simpler for both the producer and the consumer. Wait until you find the need for it. Plus, if that happens, you'll have a better understanding of the actual problem you need to solve and better solutions may reveal themselves.
In very simple cases, like a single table sql query, absolutely - databases effectively have to compute the full result, sort it, and return a window. There's almost no reason to paginate here, at an API level, unless the consumer wants only a subset (say, bandwidth limitations). Sending it all at once can be a huge benefit for those that will use it all, it's both simpler and faster for all parties.
But in most real-world cases, there are at least two additional details that can add significant response time: joins (when not involved in sorting) and additional data-gathering needed to fully build the response (e.g. getting data from other systems, internal or external). Joined data is not typically loaded prior to computing limit/offset since it may be a massive waste, and external data is effectively the same issue but with far higher latency.
And that's before getting into other practical issues, e.g. systems that can't process the response stream as it comes in - a subset will load-and-return faster than the whole content in all cases, so e.g. a website loading some json can show initial UI faster while loading more in the background. Streaming is often possible and that'll negate a lot of the downsides, but it's far less common than processing a request only after it completes.
Speaking from experience...we want to make it easy but also want to keep it performant. Getting the data all in one go is generally not performant and is easy to abuse as an API consumer. For example, always asking for all of the data rather than maintaining a cursor and secondary index (which is so much more performant for everyone involved).
Is there any good literature or patterns on supporting dumps in the tens of millions or larger?
I wrote a sheets plug-in that uses our cursor API to provide a full dump into a spreadsheet. Our professional services team is in love with it, so I bet they'd love generic data export capability.
The two main things you need are: 1. HTTP is a streaming protocol. You don't need to fully manifest a response in memory before you send it. If your framework forces that, bypass it for this particular call. (If you can't bypass it... your framework choice is now a problem for you.)
2. You presumably have some sort of JSON encoder in your language. As long as it doesn't have some sort of hard-coded "close the stream once we send this JSON" behavior (and if so, file a bug because a JSON encoder has no business doing that), all you have to do is ensure that the right bytes go out on the wire. You, again, don't have to fully manifest the reply in memory before you encode it. Something like:
stream.Write("[")
needComma = False
for item in toBeSerialized:
if needComma:
stream.Write(",")
json.Write(stream, item)
needComma = True
stream.Write("]")
A lot of times when you're emitting gigabytes of JSON it's still in lots of little chunks where each individual chunk isn't a memory problem on its own, so doing something like this can be very memory-efficient, especially if "toBeSerialized" is itself something like a cursor coming out of a DB where it itself is not manifesting in memory. (Newlines are also a good idea if your JSON encoder isn't naturally doing it already; helps debugging a lot for very little cost.)JSON objects can be more annoying; you may need to manually deserialize one and that's more annoying. Protip: Whenever possible, use the JSON encoder in your language; there is no shame or anything in using the JSON encoder to emit strings corresponding to the keys of your object. Much like HTML, you need to be very careful writing things directly to the stream; it really always should go through the encoder. I even send constant strings through the encoder just to make the code look right.
The last little tidbit is that the HTTP software stack will tend to fight you on the matter of keeping long-lived connections open. There can be a lot of places that have timeouts you may want to extend. If this gets too big you may need to do something other than HTTP. You may also need to consider detecting failures (hashing or something) and the ability to restart. (Although don't underestimate modern bandwidth and the speed you can iterate through SELECT -type queries; definitely check into the virtues of just retrying. 10GB/year of extra bandwidth and processing power is still cheaper than a developer even designing* a solution to that problem, let alone implementing and testing it.)
Oh, and if you can use HTTP, be sure you're gzip'ing. It's dirt cheap nowadays on the CPU; only in extreme situations of bandwidth abundance and CPU shortage can it be worth skipping. My rule-of-thumb on JSON shrinking is about 15:1. CSVs don't quite shrink that much but they still shrink down pretty well.
Also back in the day IBM had XML for Logging format, where every separate line was an XML fragment [2].
Most markup languages suffering of having a root element, which prevents efficient logging or steaming.
This can also be helpful when you're outputting, say, a 10MB JSON object, which isn't necessarily that large anymore, on a server where you'd like to not have to allocate 10MB of RAM to it. You can stream out a plain ol' JSON object without the resource usage.
Sadly, it is indeed easier to convince environments to stream JSON out than it is to stream it in.
Edit to add: this isn’t some wild idea, recent network and web standards have made increasing concurrency both automatic (HTTP/2 and beyond) and tunable (various prefetch APIs).