Counterpoint - CSV is absolutely atrocious and should be cast into the Sun.

It's unkillable, like many eldritch horrors.

> The specification of CSV holds in its title: "comma separated values". Okay, it's a lie, but still, the specification holds in a tweet and can be explained to anybody in seconds: commas separate values, new lines separate rows. Now quote values containing commas and line breaks, double your quotes, and that's it. This is so simple you might even invent it yourself without knowing it already exists while learning how to program.

Except that's just one way people do it. It's not universal and so you cannot take arbitrary CSV files in and parse them like this. You can't take a CSV file constructed like this and pass it into any CSV accepting program - many will totally break.

> Of course it does not mean you should not use a dedicated CSV parser/writer because you will mess something up.

Yes, implementers often have.

> No one owns CSV. It has no real specification

Yep. So all these monstrosities in the real world are all... maybe valid? Lots of totally broken CSV files can be parsed as CSV but the result is wrong. Sometimes subtly.

> This means, by extension, that it can both be read and edited by humans directly, somehow.

One of the very common ways they get completely fucked up, yes. Someone goes and sorts some rows and boom broken, often unrecoverable data loss. Someone doesn't correctly add or remove a comma. Someone mixes two files that actually have differently encoded text.

> CSV can be read row by row very easily without requiring more memory than what is needed to fit a single row.

CSV must be parsed row by row.

> By comparison, column-oriented data formats such as parquet are not able to stream files row by row without requiring you to jump here and there in the file or to buffer the memory cleverly so you don't tank read performance.

Sort of? Yes if you're building your own parser but who is doing that? It's also not hard with things like parquet.

> But of course, CSV is terrible if you are only interested in specific columns because you will indeed need to read all of a row only to access the part you are interested in.

Or if you're interested in a specific row, because you're going to have to be careful about parsing out every row until you get there.

CSV does not have a row separator. Or rather it does but it also lets you have that row separator appear and not mean "separate these rows" so you can't simply trust it.

> But critics of CSV coming from this set of pratices tend to only care about use-cases where everything is expected to fit into memory.

Parquet uses row groups which means you can stream chunks easily, those chunks contain metadata so you can easily filter rows you don't need too.

I much more often need to keep the whole thing in memory working with CSV than parquet. With parquet I don't even need to be able to fit all the rows on disk I can read the chunk I want remotely.

> CSV can be appended to

Yeah that's easier. Row groups means you can still do this though, but granted it's not as easy. *However* I will point out that absolutely nothing stops someone completely borking things by appending something that's not exactly the right format.

> CSV is dynamically typed

Not really. Everything is strings. You can do that with anything else if you want to. JSON can have numbers of any size if you just store them as strings.

> CSV is succinct

Yes, more so than jsonl, but not really more than (you guessed it) parquet. Also it's horrific for compression.

> Reverse CSV is still valid CSV

Get a file format that doesn't absolutely suck and you can parse things in reverse if you want. More usefully you can parse just sections you actually care about!

> Excel hates CSV

Helpfully this just means that the most popular way of working with tabular data in the world doesn't play that nicely with it.

I agree with all this and will just add. CSV is slow. Like really slow.

Partially due to the bloat, but also partially because the format doesn't allow for speed.

And because CSV is untyped, you have to either trust the producer or put in mountains of guards to ensure you can handle the weird garbage that might come through.

My company deals with a lot of CSV and we literally built tools and hire multiple full time employees whose entire job is handling CSV sucking in new and interesting ways.

Parquet literally eliminates 1/2 of our our data ingestion pipeline simply by being typed, consistent, and fast to query.

One example of a problem we constantly run into is that nobody likes to format numbers the same way. Scientific notation, no scientific notation, commas or periods, sometimes mixed formats (scientific notation when a number is big enough, for example).

Escaping is also all over the board.

CSV SEEMS simple, but the lack of a real standard means it's anything but.

I'd take xml over CSV.

Or just forget about the quotes which open a new can of worms, and use TSV while escaping newlines and tabs in values.

Can you provide a reproducible example of how sorting rows can lead to unrecoverable data loss?

Also, commas in quoted strings are quite mainstream csv, but csvs with quoted strings containing unescaped newlines are extremely baroque. Criticism of csv based on the assumption that strings will contain newlines is not realistic.