So, streaming one new record in and seeing how this changes the results of a multi-way join with many other large relations can happen in milliseconds in TD, vs batch systems which will re-read the large inputs as well.

This isn't a fundamentally new difference; Flink had this difference from Spark as far back as 2014. There are other differences between Flink and TD that have to do with state sharing and iteration, but I'd crack open the papers and check out the obligatory "related work" sections each should have.

For example, here's the first para of the Related Work section from the Naiad paper:

> Dataflow Recent systems such as CIEL [30], Spark [42], Spark Streaming [43], and Optimus [19] extend acyclic batch dataflow [15, 18] to allow dynamic modification of the dataflow graph, and thus support iteration and incremental computation without adding cycles to the dataflow. By adopting a batch-computation model, these systems inherit powerful existing techniques including fault tolerance with parallel recovery; in exchange each requires centralized modifications to the dataflow graph, which introduce substantial overhead that Naiad avoids. For example, Spark Streaming can process incremental updates in around one second, while in Section 6 we show that Naiad can iterate and perform incremental updates in tens of milliseconds.

If you can keep a hash map of the things you've seen, then it is easy to respond quickly to that one new thing. If you are not allowed to maintain any state, then you don't have a lot of options to efficiently respond to the new thing, and most likely need to re-read the 1M things.

That's the benefit of being stateful. There is a cost too, which is that you need to be able to reconstruct your state in the case of a failure, but fortunately things like differential dataflow (built on TD) are effectively deterministic.

Also, I suspect "Spark" is a moving target. The original paper described something that was very much a batch processor; they've been trying to fix that since, and perhaps they've made some progress in the intervening years.

I think Spark does that sometimes these days, but I don't know much about the specifics of how and when Spark does it.

Does TD have to keep _everything_ in memory, or can it be strategic in what it keeps and what it evicts?

Differential dataflow plugs in certain logic there, and it does indeed maintain a synopsis of what data have gone past, sufficient to respond to future updates but not necessarily the entirety of data that it has seen.

It would be tricky to implement DD over classic Spark, as DD relies on these synopses for its performance. There are some changes to Spark proposed in recent papers where it can pull in immutable LSM layers w/o reading them (e.g. just mmapping them) that might improve things, but until that happens there will be a gap.