Hacker News Reader

Comment by busterarm | Hacker News Reader

This isn't meant to offend, rather as a point of consideration, but seeing your example use case being 10GB and then talk about big data frameworks makes it hard for me to take this advice seriously.

I might reach for that kind of tooling at the hundreds of TB to PB scale, but in our production applications we have _tables_ that are multiple terabytes. SQL is just fine.

Yes, we also have have queries that run in the timescale of hours and they are always of the reporting/scheduled task variety and absolutely vital to our customers. Long running reporting queries are pretty acceptable (and pretty much the norm since forever) outside of the tech industry and your customers won't balk at it.

View on HN

It seems like you misunderstood what I wrote. I'm saying you should consider using R or Python if your reports are taking a long time, not big data frameworks.

Big data was a reference to thinking about the problem in terms of the speed of the hardware. If it's 1000x slower than what the hardware can do, that's a sign you're using the wrong tool for the job.

Getting within 10x is reasonable, but not 100x or 1000x, which is VERY COMMON in my experience. These two situations are very common:

1) SQL queries that are orders of magnitude slower than a simple offline computation in Python or R (let alone C++). The underlying cause is usually due to bad query planning of joins / lack of indices.

You might not have the ability to add indices easily, and even if you did, that has its drawbacks for one-off queries.

2) You need to do some computation that's awkward inside SQL. Statistics beyond basic aggregations, iterative computations (loops), and tree structures are common problems.

ETL pipelines that hop between different kinds of storage/computing platforms to exploit local maxima, like you're pointing out with R and SQL working in concert, is pretty common in companies working their way up to BigData and academia.

From the Enterprise side I think too many developers have an unfounded expectations around data storage technology. There's this unchallenged belief that monolithic datastorage that will solve thier problems across the entire time/storage/complexity spectrum. By bringing multiple tools to bear, instead, you end up with more purpose built storage but far less domain impedence.

Slapping a denormalized NoSQL front-end for webscale onto a legacy RDBMS can be a cheap win/win to maximize the capabilities of both. SQL + R is oodles better than R or SQL in isolation.

Reply on news.ycombinator.com