The Dataflow Model Revisited

Posted by scott_s 1 day ago

Counter105Comment14OpenOriginal

PDF: https://www.vldb.org/pvldb/vol19/p4953-fernandez-moctezuma.p...

Comments

Comment by scott_s 1 day ago

I worked in the streaming area for a decade, doing research and development (see: https://scholar.google.com/citations?user=Rdf5OIYAAAAJ&hl=en). After moving on from streaming specifically and moving into the general problems in large data warehouses, I also concluded: just default to SQL for all analytics and the database lens is the best way to think about streaming for analytics.

I still do think that stream programming models are extremely interesting and powerful. But I used to think they would eventually become more mainstream as a way to elegantly program for high throughput, low latency massively parallel systems. That has not been the case, and I no longer think that it will be. People get by with the existing programming languages and models, that seems to be fine.

Comment by necubi 21 hours ago

I've similarly spent a decade in the streaming space, started a stream processing startup, built three streaming platforms at various large companies... and I basically agree.

Streaming always felt like it was about to happen, and it kept not happening, in a year of the linux desktop sort of way. I (and others, who also optimistically started stream processing startups) thought it was a technology problem but ultimately it's a demand problem: very few companies actually need low latency processing. And continuously running, stateful systems are inherently more complex to operate and evolve compared to batch systems.

Now that relatively low-latency ingest into warehouses and datalakes is easily achieved, it's really hard to make the argument to invest in complicated streaming systems.

That said, it still feels like we've landed in a suboptimal spot. Stream processors (in particular, those following the dataflow model) solve two problems that are hard in batch systems:

1. Determining and signaling completeness for a time period (when have I likely received all of the data for 12:01pm, such that I can now safely process it)

2. Avoiding expensive recomputations for periodic queries

I do think long term we'll see some amount of streaming influence moving back into the batch systems, particularly as object storage gives us the ability to run stateful workloads with less operational headache.

(If anyone else finds these problems interesting, I'm hiring for my stream processing team at Cloudflare. Email in my profile)

Comment by scott_s 20 hours ago

Agreed agreed.

During my time in the space, my pithy saying about the system I worked on [1] was that we could scale up or down. If you wanted to do streaming packet filtering with microsecond latency, we could do that. If you wanted to do complex analytics on structured data, we could do that. We did have deployments that "scaled down" and were more stream processing rather than streaming analytics. But analytics is by far the dominant use case, and SQL and relational databases are the better abstraction there. And for the stream processing cases, folks tend to stick to their existing lower-level stacks.

[1] I worked on IBM Streams, https://www.ibm.com/docs/en/streams/4.3.0?topic=welcome-intr..., which had its own language, compiler and runtime system. IBM sold this technology in 2023: https://21cs.com/en/resources/articles/2023/10/10/21cs-acqui....

Comment by convolvatron 20 hours ago

the fact that traditional database infrastructure handles many people's needs doesn't really obviate dataflow style architectures. personally I think (3) composition is a real potential win. I would also note that streaming is a really great base on which to build a distributed database that looks more like Postgres on top.

I guess I'm mostly confused about the idea that streaming systems should be trying to supplant sql databases, or that a failure to do so implies that they don't have utility. it looks like the GP has done some really interesting work on automatic parallelism. that isn't pointless just because most people building operational systems still reflexively reach for PG.

Comment by scott_s 19 hours ago

That work is still relevant! It's just that end-users don't need to be aware of it. The position of the paper I submitted, which I basically agree with, is that "streaming" shouldn't need to be something end-users care about. It's something the system does based on needs.

Databases already have a dataflow style architecture: that's how they implement queries. Because SQL is relational, SQL queries become dataflow execution plans.

One way to think about the programming model I worked on is that it was like exposing a query plan API directly to users, instead of giving them SQL.

Comment by zekrioca 16 hours ago

You paper "A catalog of stream processing optimizations" is great! It seems like a classic in stream processing for me :)

Generally I agree with this comment. Very few niche applications require stream processing and all struggle that had been into research to enable some of the ACID properties for stream processing.

While at it, would you know how kdb+ (mostly used in finance) work? It -- together with q -- seems to achieve very low latency for real time systems.

Comment by ModernMech 21 hours ago

> People get by with the existing programming languages and models, that seems to be fine.

People like what they like, and they have trouble with anything other than imperative structured programming. It’s been a pain over the last decade plus to get people just to try functional programming, and they already know what functions are!

It’s so hard I think it’s a losing battle. No matter how good a paradigm people will always stick to a) what they know and b) what they’re paid to know.

However, I still think there’s a future for paradigms like dataflow because LLMs use random esoteric languages with great ease, and without complaint. They will use the tool for the job, even if they’re unfamiliar with it at first. “Mainstream” doesn’t really matter anymore.

Comment by porridgeraisin 21 hours ago

There is a reason imperative programming is so common. It is more amenable to poorly designed, under specified, iterative development. Most real world software is in that category naturally. If you're meticulously designing and engineering it really well from the start, sure functional languages represent it well without leaving much room for misinterpretation and thus bugs. But no one does that.

Marginal cost of adding a feature has to be proportional to the revenue made by that feature. Then In Java or go you just add a ugly special case to appease the large customer and ignore the small ones' emails. Bugs getting shunted around instead of truly fixed at the root is also totally OK as long as they are not in the major revenue/cost centers of the product. No one has time to replace these piles of hacks and eventually you would have given up most of your languages benefits and your types now mean nothing there is probably a hundred flags making it a union effectively.

Rust is one language though where hacking around goes a long way without breaking too many of the guarantees, although it's not perfect, from my experience at a company where services spanned java go rust and ruby.

Comment by BenoitP 21 hours ago

> That has not been the case, and I no longer think that it will be. People get by with the existing programming languages and models, that seems to be fine.

.. as long as the processing they do fits in a map().

I still deeply believe that a new class of programs can be unlocked by using reduce() or all_gather() in a massive concurrent stream.

In a stadium, why can't a uuid be embedded as an acoustic chirp emitted by all the smartphones, then triangulation-reduced in a 3D map of all devices, enabling a stadium-wide sound system?

Why don't we have MMORPGs that can deal with 100k users in a single game instance?

Comment by janpeuker 21 hours ago

I used to be extremely into Dataflow/Apache Beam and literally had the paper printed on my desk AND the book. I agree they got Event time versus processing time and Never rely on completeness right, and I love they go deeper into why that was just a hard thing to accept. My head still hurts thinking about unbounded stream triggers and I'm happy we got around to a table-centric model. I still think taking some ideas from Spanner, databases as message bus or consistency information per row, basically in-database CQRS would have been nice, though. Great paper.

Comment by jasonwatkinspdx 19 hours ago

I've always thought this paper was a bit of an overlooked gem: https://www.vldb.org/conf/2002/S07P01.pdf

It views stream processing as a join between queries and data, where internally everything is just materialized views probing each other.

Interesting that the authors above ended up landing in a similar place conceptually.

Comment by guglecwoam 17 hours ago

> Splitting

This is one of the things I wished for in the Beam APIs, and I'm glad they included it in the doc. Some (single) elements are bigger than a single function, and can't be processed one at a time. For example. a very large CSV file. They added support for it later (as Splittable DoFn's) but each implementation had varying support for it.

Comment by 7e 18 hours ago

Streaming is expensive, both in terms of machine and human cost, and the benefit is quite marginal for most applications. However, Google engineers need to get promotion somehow, you know?

Comment by kevinbaiv 14 hours ago

[flagged]

Comment by rawgabbit 16 hours ago

TLDR. 15 years ago they avoided SQL and created a language for streaming. Today, the industry decided that SQL materialized views with a “freshness” or time lag promise is the correct mental model.

Comment by lincpa 22 hours ago

[dead]