DuckDB as the New jq | Hacker News Reader

DuckDB as the New jq

pgrs.net

364 pointspgr0ss73 comments

The most effective combination I've found so far is jq + basic shell tools.

I still think jq's syntax and data model is unbelievably elegant and powerful once you get the hang of it - but its "standard library" is unfortunately sorely lacking in many places and has some awkward design choices in others, which means that a lot of practical everyday tasks - such as aggregations or even just set membership - are a lot more complicated than they ought to be.

Luckily, what jq can do really well is bringing data of interest into a line-based text representation, which is ideal for all kinds of standard unix shell tools - so you can just use those to take over the parts of your pipeline that would be hard to do in "pure" jq.

So I think my solution to the OP's task - get all distinct OSS licenses from the project list and count usages for each one - would be:

curl ... | jq '.[].license.key' | sort | uniq -c

That's it.

> I still think jq's syntax and data model is unbelievably elegant and powerful once you get the hang of it - but its "standard library" is unfortunately sorely lacking in many places

After a few years of stalled development, jq has been taken over recently by a new team of maintainers and is rapidly working through a lot of longstanding issues (https://github.com/jqlang/jq), so I'm not sure if this is still the case

Wasn't aware of that, that's great to hear! I think if there is one utility that deserves a great maintainer team then this one. But if we saw some actual improvements in the future, that would be awesome!

I have a list of pet peeves that I'd really like to see fixed, so I'm gonna risk a bit of hope.

The Unix philosophy continues to pass the test of time.

Yes and no. Many UNIX philosophy proponents are abhorred by powerful binaries like jq and awk.

Like any religion, there are zealots and "no true Scotsman" arguments a'plenty.

> I still think jq's syntax and data model is unbelievably elegant and powerful once you get the hang of it [...]

It's basically just functional programming. (Or what you would get from a functional programmer given the task of writing such a tool as jq.)

That's not to diminish jq, it's a great tool. I love it!

As an old Unix guy this is exactly how I see jq: a gateway to a fantastic library of text processing tools. I see a lot of complicated things done inside the language, which is a valid approach. But I don’t need it to be a programming language itself, just a transform to meet my next command after the pipe.

If I want logic beyond that, then I skip the shell and write “real” software.

I personally find those both to be more readable and easier to fit in my head than long complex jq expressions. But that’s completely subjective and others may find the jq expression language easier to read than shell or (choose your programming language).

Your comment made me go look up jq (even more than the article did) and the first paragraph of the repo [0] feels like a secret club's secret language.

I'm very interested, but not a Linux person, do you know of any good resources for learning the Linux shell as a programming language?

[0] https://jqlang.github.io/jq/

I’ll say, I did shell scripting for years from copy/paste, cribbing smarter people, and reading online guides. But I didn’t really understand until I read The Unix Programming Environment by Brian Kernighan and Rob Pike.

It’s a very old book and the audience was using dumb terminals. But it made me understand why and how. I think I’ve read every Kernighan book at this point and most he was involved in because he is just so amazing and not just conveying facts, but teaching how to think idiomatically in the topic.

I also used awk for 2 decades, kind of like how I use jq now. But when I read his memoir I suddenly “got it.” What I make with it now is intentional and not just me banging on the keyboard until it works. A great middle ground for something a little sophisticated, but not worth writing a full program for.

Something else that helped me was to install a minimal distro… actually a base FreeBSD install would be great… and read the man pages for all the commands. I don’t remember the details, but I learned that things existed. I have many man pages that I look at the same options on every few months because I’m not positive I remember right. Heck, I ‘man test’ all the time still. (‘test’ and ‘[‘ are the same thing)

I also had an advantage of 2 great coworkers. They’d been working on Unix since the 80s and their feedback helped me be more efficient, clean, and avoid “useless use of cat” problems.

I also highly recommend using shellcheck. I sometimes disagree with it when I’m intentionally abusing shell behavior, but it’s a great way to train good habits and prevent bugs that only crop up with bad input, scale, etc. I get new devs to use it and it’s helped them “ramp up” quickly, with me explaining the “why” from time to time.

But yeah. The biggest problem I see is that people think there is more syntax than there really is (like my test and [ comment). And remember it’s all text, processes, and files. Except when we pretend it’s not ;).

> The Unix Programming Environment

How does this compare to The Art of Unix Programming, if you've read both?

I don’t find that book to be very useful at all.

I’m kind of annoyed by the bait and switch of the title. It’s a play on Knuth’s classic but then turns into showing why Unix/Linux is better than Windows, etc.

As a disclaimer: I really don’t respect ESR and his work, and admire Brian Kernighan immensely. Very odd to be in a situation where those names are put side by side. Just want to call out that I do have bias on the people here. Don’t want to get into why as that’s not constructive.

I wasn't aware of the bait and switch at the time I read it, but I did really enjoy the history of how the Unix/Linux ethic came together and evolved over time. Had I heard of The Unix Programming Environment when I read it in 2014 I may have gone with that instead, as I was looking for something more along the lines of a technical handbook rather than a code of ethics.

Yeah and ESR can be revisionist in his history, projecting intention on something organic. He alienated a lot of people over time with this… and other behavior.

The book I recommended is both a handbook and a “how to think.” It applies forward to things introduced well after the book. But it also helped me understand why the Byzantine behavior of a tty is what it is.

If you are interested in the history from a first person perspective, I do recommend Kernighan’s “Unix: A History and a Memoir”. He went from originally trying to write something objective to realizing it was necessarily his personal experience. Even the culture aspect of his story has influenced how I try to foster teamwork. It was an engaging read for me.

I felt it was a good dive into "good" programs at a bit of higher level, rather than "here's how to do X". Quite a bit applies to Windows software and other software that never touches Unix as well.

Some bits are better than others, some bits haven't aged too well in the last 20 years, and it's a shame esr has since turned crazy. But I still feel the book holds up reasonably well.

"Bait and switch" certainly seems too strong of an accusation, especially for a book that's available for free.

I do agree that even pre-crazy esr was never on the level of Kernighan in any way.

Really love your comment, so much that I wanted to check out the books you mentioned.

After searching z-lib for "The UNIX Programming Environment", all I found was a janky and grainy PDF. Then I searched archive.org and discovered this high fidelity PDF version:

https://archive.org/details/UnixProgrammingEnviornment

Note: Sadly, the EPUB version is 10x larger (370MB) and is corrupted, not able to be opened / viewed.

Much thanks for this. I've been struggling to grasp what's happening down in the engine room beneath me as I operate a modern Linux environment and only knew that there have been different waves of evolutions over the decades. I instead only see today's pretty deck up top, outside.

I'm only ten pages in, but that's mostly because the format and approach of this book is quickly yielding me tons of conversation fodder for ChatGPT4, where I've been endlessly asking it to clarify or point out holes in my mental model of everything I'm doing in the terminal, and still working out things such as why Backspace, Ctrl+H, Ctrl+/, Ctrl+?, Ctrl+-, and Ctrl+_ all together seem to have some overlap or differences in what's happening, depending on terminal contexts. I always had working notions of raw versus cooked sessions, the line discipline, etc., but I'm finding ways to play with the machinery with silly exercises that I otherwise couldn't have come up with.

For example, I just opened up three terminal windows:

  1) `man ascii`
  2) `nc -lvp 9001 | xxd -c1`
  3) `stty raw -echo; nc -nv 127.0.0.1 9001`

Then, reproducing the whole "Hex" column of the manpage, in order, using my keystrokes. And, observing the multi-byte payloads of some of the other keys. I feel like a kid again :)

So, grab yourself a Linux box (I suggest Debian), a large CSV file or JSON lines file you need to slice up, and an hour of time, and start trying out some bash one-liners on your data. Set some goals like "find the Yahoo email addresses in the data and sort by frequency" or "find error messages that look like X" or "find how many times Ben Franklin mentions his wife in his autobiography"

Here's the thing. These tools have been used since the '70s to slice, dice and filter log files, CSVs, or other semi-structured data. They can be chained together with the pipe command. Sys admins were going through 100MB logs with these tools before CPUs hit the gigahertz

These tools are blisteringly fast, and they are basically installed on every Linux machine.

https://github.com/onceupon/Bash-Oneliner

And for a different play-by-play example:

https://adamdrake.com/command-line-tools-can-be-235x-faster-...

>Your comment made me go look up jq (even more than the article did) and the first paragraph of the repo [0] feels like a secret club's secret language.

Or one of the most old standing widespread clubs of computing open standard language :)

"jq is like sed for JSON data - you can use it to slice and filter and map and transform structured data with the same ease that sed, awk, grep and friends let you play with text."

Translation:

JQ is like a (UNIX/POSIX staple command line text-manipulation tool) but specialized for text structured in JSON format. You can use it to extract parts of a JSON document (slice), keep nodes based on some criteria (filter), transform each element in a list of structured data to get a new list with the transformed versions (map), and do that as easily as you can with the sed (basic command line text manipulation program), awk (command line text manipulation program with a full featured text-processing oriented language), grep (command line program to search for strings), and other assorted unix userland programs.

Your command line solution doesn't give quite the same result as OP. The final output in OP is sorted by the count field, but your command line incantation doesn't do that. One might respond that all you need to do is add a second "| sort" at the end, but that doesn't quite do it either. That will use string sorting instead of proper numeric sorting. In this example with only three output rows it's not an issue. But with larger amounts of data it will become a problem.

Your fundamental point about the power of basic shell tools is still completely valid. But if I could attempt to summarize OP's point, I think it would be that SQL is more powerful than ad-hoc jq incantations. And in this case, I tend to agree with OP. I've made substantial use of jq and yq over the course of years, as well as other tools for CSVs and other data formats. But every time I reach for them I have to spend a lot of time hunting the docs for just the right syntax to attack my specific problem. I know jq's paradigm draws from functional programming concepts and I have plenty of personal experience with functional programming, but the syntax and still feel very ad hoc and clunky.

Modern OLAP DB tools like duckdb, clickhouse, etc that provide really nice ways to get all kinds of data formats into and out of a SQL environment seem dramatically more powerful to me. Then when you add the power of all the basic shell tools on top of that, I think you get a much more powerful combination.

I like this example from the clickhouse-local documentation:

  $ ps aux | tail -n +2 | awk '{ printf("%s\t%s\n", $1, $4) }' \
      | clickhouse-local --structure "user String, mem Float64" \
          --query "SELECT user, round(sum(mem), 2) as memTotal
            FROM table GROUP BY user ORDER BY memTotal DESC FORMAT Pretty"

For reference, sort(1) has -n, --numeric-sort: compare according to string numerical value.

You can archive that by appending sort -n, so the whole thing becomes:

curl ... | jq '.[].license.key' | sort | uniq -c | sort -n

You can even turn it back into json by exploiting the fact that when uniq -c gets lines of json as input, it's output will be "accidentally" parseable as a sequence of json literals by jq, where every second literal is a count. You can use jq's (very weird) input function to transform each pair of literals into a "proper" json object:

curl ... | jq '.[].license.key' | sort | uniq -c | sort -n | jq '{"count":., "value":input}'

> I think it would be that SQL is more powerful than ad-hoc jq incantations.

I don't disagree, but only because I have a lot of experience with SQL so I "think" more in those terms. The blog post made perfect sense to me, and I have to go to claude/openai/copilot for whatever I need in `jq`, EVERY TIME. Because I don't use it nearly as often, so I don't have its language internalized.

Readability (and I'd posit to the point of this, power) is more a function of the reader than the code, pg's "blub paradox" notwithstanding.

I didn't know anything about duckdb, but I might give this a shot.

found out recently that jq can url-encode values, is there anything it _can't_ do?

Finding out whether . is contained in a given array or not, evidently.

(That's not strictly true - you can do it, you just have to bend over backwards for what is essentially the "in" keyword in python, sql, etc. jq has no less than four functions that look like they should do that - in(), has(), contains() and inside(), yet they all do something slightly different)

Driving that further - I don’t want to have to edit my query for minutes when I am in the shell. I don’t believe that the shell is conducive to complex SQL queries. You could write simpler SQL queries, but then you’re in a space where there are less verbose tools for those simpler tasks.

DuckDB seems stronger for someone who needs to create a scripting library - which also has lots of options and competition - or someone who has a very specific workflow of working with JSON dumps for a huge percentage of their time.

I do the same thing all the time, lol

but have in the last year or so, tried to start writing things that will be read/used by other people in python or java

why? because most people don't have a clue how bash scripts work, but can read/debug python and java with ease, and both can be run as shell scripts too

the mac has sqlite installed by default as well, so there's a powerful combo available on every mac

but I totally do what you're doing all the time, and it is my default go to if I have to get something done fast :D

If you like lisp, and especially clojure, check out babashka[0]. This my first attempt but I bet you can do something nicer even if you keep forcing yourself to stay into a single pipe command.

  cat repos.json | bb -e ' (->> (-> *in* slurp (json/parse-string true))
                                (group-by #(-> % :license :key))
                                (map #(-> {:license (key %)
                                           :count (-> % val count)}))
                                json/generate-string
                                println)'

[0] https://babashka.org/

Related, clickhouse local cli command is a speed demon to parse and query JSON and other formats such as CSV:

- "The world’s fastest tool for querying JSON files" https://clickhouse.com/blog/worlds-fastest-json-querying-too...

- "Show HN: ClickHouse-local – a small tool for serverless data analytics" https://news.ycombinator.com/item?id=34265206

I'll second this. Clickhouse is amazing. I was actually using it today to query some CSV files. I had to refresh my memory on the syntax so if anyone is interested:

  clickhouse local -q "SELECT foo, sum(bar) FROM file('foobar.csv', CSV) GROUP BY foo FORMAT Pretty"

Way easier than opening in Excel and creating a pivot table which was my previous workflow.

Here's a list of the different input and output formats that it supports.

https://clickhouse.com/docs/en/interfaces/formats

You don't even need to use file() for a lot of things recently. These just work with clickhouse local. Even wildcards work.

  select * from `foobar.csv`

  select * from `monthly-report-*.csv`

Ooh very nice, thanks for the tip!

Just had to try:

  $ function _select_aux () { clickhouse local -q "SELECT $* FORMAT Pretty" }
  $ alias SELECT='noglob _select_aux'
  $ SELECT COUNT(*) as count FROM file('repos.json', JSON)
  ┏━━━━━━━┓
  ┃ count ┃
  ┡━━━━━━━┩
  │    30 │
  └───────┘

This is very nice!! I went down the format clause[1] rabbit hole a couple minutes ago and it looks like they have `FORMAT Markdown` as an option. FORMAT PrettySpaceNoEscapes also looks a bit better than Pretty or the default PrettyCompact imho

1. https://clickhouse.com/docs/en/interfaces/formats

Where's this noglob command coming from? I haven't heard of it before.

Ah sorry that is probably zsh specific, it makes sure to not glob expand * etc

Very cool!

I am also a big fan of jq.

And I think using DuckDB and SQL probably makes a lot of sense in a lot of cases.

But I think the examples are very geared towards being better solved in SQL.

The ideal jq examples are combinations of filter (select), map (map) and concat (.[]).

For example, finding the right download link:

  $ curl -s https://api.github.com/repos/go-gitea/gitea/releases/latest \
    | jq -r '.assets[]
             | .browser_download_url
             | select(endswith("linux-amd64"))'
  https://github.com/go-gitea/gitea/releases/download/v1.15.7/gitea-1.15.7-linux-amd64

Or extracting the KUBE_CONFIG of a DigitalOcean Kubernetes cluster from Terraform state:

  $ jq -r '.resources[]
          | select(.type == "digitalocean_kubernetes_cluster")
          | .instances[].attributes.kube_config[].raw_config' \ 
      terraform.tfstate
  apiVersion: v1
  kind: Config
  clusters:
  - cluster:
      certificate-authority-data: ...
      server: https://...k8s.ondigitalocean.com
  ...

I think that's a fair point. Unnesting arrays in SQL can be annoying. Here is your first example with duckdb:

  duckdb -c \
    "select * from ( \
      select unnest(assets)->>'browser_download_url' as url \
      from read_json('https://api.github.com/repos/go-gitea/gitea/releases/latest') \
    ) \
    where url like '%linux-amd64'"

In case someone else was wondering, one can get a shell-consumable output from that with

  duckdb -noheader -list

I tried this and it just seems to add bondage and discipline that I don't need on top of what is, in practice, an extremely chaotic format.

Example: trying to pick one field out of 20000 large JSON files that represent local property records.

% duckdb -json -c "select apn.apnNumber from read_json('*')" Invalid Input Error: JSON transform error in file "052136400500", in record/value 1: Could not convert string 'fb1b1e68-89ee-11ea-bc55-0242ad1302303' to INT128

Well, I didn't want that converted. I just want to ignore it. This has been my experience overall. DuckDB is great if there is a logical schema, not as good as jq when the corpus is just data soup.

You can also query (public) Google Sheets [0]

    SELECT * 
    FROM read_csv_auto('https://docs.google.com/spreadsheets/export? 
    format=csv&id=1GuEPkwjdICgJ31Ji3iUoarirZNDbPxQj_kf7fd4h4Ro', normalize_names=True);

0 - https://x.com/thisritchie/status/1767922982046015840?s=20

In a similar vein, I have found Benthos to be an incredible swiss-army-knife for transforming data and shoving it either into (or out of) a message bus, webhook, or a database.

https://www.benthos.dev/

How does this defer from filebeat?

I don't know which filebeat you are referring to...

https://github.com/elastic/beats/tree/master/filebeat

This one? I only looked for a moment, but filebeat appears to be ingestion only. Benthos does input, output, side-effects, stream-stream joins, metrics-on-the-side, tiny-json-wrangling-webooks, and more. I find it to be like plumbers putty, closing over tooling gaps and smoothing rough edges where ordinarily you'd have to write 20 lines of stream processing code and 300 lines of error handling, reporting, and performance hacks.

I wish it was not based on YAML. Pipelines are code, not configuration!!

I've found Nushell (https://www.nushell.sh/) to be really handy for ad-hoc data manipulation (and a decent enough general-purpose shell).

Nushell is really good but the learning curve is massive.

I've been on nushell for almost a year now and still struggle to put more complex commands together. The docs are huge but not very good and the community resources are very limited (it's on Dicord smh) unfortunately. So, if anyone wants to get into it you really need to put down few days to understand the whole syntax suite but it's worth it!

I run a pretty substantial platform where I implemented structured logging to SQLite databases. Each log event is stored as a JSON object in a row. A separate database is kept for each day. Daily log files are about 35GB, so that's quite a lot of data to go through is you want to look for something specific. Being able to index on specific fields, as well as express searches as SQL queries is a real game changer IMO.

I have a lot of trouble understanding the benefits of this versus just working with json with a programming language. It seems like you're adding another layer of abstraction versus just dealing with a normal hashmap-like data structure in your language of choice.

If you want to work with it interactively, you could use a notebook or REPL.

if you are used to the command line and knows some basic syntax, it is less verbose then opening a REPL and reading a file. The fact that you can pipe the json data into it is also a plus, making it easier to check quickly if the response of a curl call has the fields/values you were expecting. Of course, if you are more comfortable doing that from the REPL, you get less value from learning jq. If you are fond of one liners, jq offers a lot of potential.

Pipelining CLI commands or bash scripts. From a security perspective, it may be preferable to not ship with a runtime.

bash and jq are both runtimes.

Very difficult (or often impractical) to not have a shell at all, and jq is at least limited in scope, and has no dependencies that need to be installed. Far better than a full language with its own standard library and set of dependencies to lock down.

Use a compiled language like golang if you don't want to ship with a runtime.

If you're willing to ship w/ bash then I don't understand the opposition to JS. Either tool puts you in a scenario where somebody who can exec into your env can do whatever they want

My thoughts as well:

  const response = await fetch("https://api.github.com/orgs/golang/repos");
  const repos = await response.json();

  const groups = Map.groupBy(repos, e => e?.license?.key);
  ...

i've been using simonw's sqlite-utils (https://sqlite-utils.datasette.io/en/stable/) for this sort of thing; given structured json or jsonl, you can throw data at an in-memory sqlite database and query away: https://sqlite-utils.datasette.io/en/stable/cli.html#queryin...

Thank you was going to say this as well, sqlite does this the same as duckdb

Worth noting that both jq and duckdb can be used from python and from the command line. Both are very useful data tools!

While jq’s syntax can be hard to remember, ChatGTP does an excellent job generating jq from an example json file and a description of how you want it parsed.

I work primarily in projects that use js and I mostly don't see the point in working with json in other tools than js.

I have tried jq a little bit, but learning jq is learning a new thing, which is healthy, but it also requires time and energy, which is not always available.

When I want to munge some json I use js... because that is what js in innately good at and it's what I already know. A little js script that does stdin/file read and then JSON.parse, and then map and filter some stuff, and at the end JSON.stringify to stdout/file does the job 100% of the time in my experience.

And I can use a debugger or put in console logs when I want to debug. I don't know how to debug jq or sql, so when I'm stuck I end up going for js which I can debug.

Are there js developers who reach for jq when you are already familiar with js? Is it because you are already strong in bash and terminal usage? I think I get why you would want to use sql if you are already experienced in sql. Sql is common and made for data munging. Jq however is a new dsl when I don't see the limitation of existing js or sql.

I do quite a lot of adhoc/exploratory programming to query and transform data then jq is very convenient as it works very well with "deep" data structures and the language itself it very composable.

To debug in jq you can use the debug function to prints to stderr, ex: "123 | debug | ..." or "{a:123, b:456} | debug({a}) | ... " only prints value of a "{a:123}"

Hi,

I very much share your sentiment and I saw a few comments mentioning PRQL so I thought it might be worth bringing up the following:

In order to make working with data at the terminal as easy and fun as possible, some time ago I created pq (prql-query) which leverages DuckDB, DataFusion and PRQL.

Unfortunately I am currently not in a position to maintain it so the repo is archived but if someone wanted to help out and collaborate we could change that.

It doesn't have much in the way of json functions out-of-the-box but in PRQL it's easy to wrap the DuckDB functions for that and with the new PRQL module system it will soon also become possible to share those. If you look through my HN comment history I did provide a JSON example before.

Anyway, you can take a look at the repo here: https://github.com/PRQL/prql-query

If interested, you can get in touch with me via Github or the PRQL Discord. I'm @snth on both.

Jq tip: Instead of `sort_by(.count) | reverse`, you can do `sort_by(-.count)`

only if you're sure that .count is never null:

  $ echo '[{"a": {"count": null}}]' | jq -c 'sort_by(-.count)'
  jq: error (at <stdin>:1): null (null) cannot be negated
  $ echo '[{"a": {"count": null}}]' | jq -c 'sort_by(.count) | reverse'
  [{"a":{"count":null}}]

this whole thread is like nerd sniping me :-D but I felt compelled to draw attention to jq's coalesce operator because I only recently learned about it and searching for the word "coalesce" in the man page is pfffft (it's official name is "Alternative operator", with alternative being "for false and null")

  $ echo '[{"a": {"count": null}}]' | jq -c 'sort_by(-(.count//0))'
  [{"a":{"count":null}}]

https://github.com/mikefarah/yq

Yq handles almost every format, and IMO easier to use.

Shoutout to jqp, an interactive jq explorer.

https://github.com/noahgorstein/jqp

This is pretty nice!

I love jq and yq, but sometimes I don't want to invest time in learning new syntax and just fallback to some python one liner, that can if necessary become a small python script.

Something like this, I have a version of this in a shell alias:

  python3 -c "import json,sys;d=json.load(sys.stdin);print(doStuff(d['path']['etc']))"

Pretty print is done with json.dumps.

There is also a way to import a table from the STDIN (see also https://duckdb.org/docs/data/json/overview)

cat my.json | duckdb -c "CREATE TABLE mytbl AS SELECT * FROM read_json_auto('/dev/stdin'); SELECT ... FROM mytbl"

DuckDB parses JSON using yyjson internally .

https://github.com/ibireme/yyjson

My current team produces a CLI binary that is available on every build system and everybody's dev machines

Whenever we're writing automation, if the code is nontrivial, or if it starts to include dependencies, we move the code into the CLI tool.

The reason we like this is that we don't want to have to version control tools like duckdb across every dev machine and every build system that might run this script. We build and version control a single binary and it makes life simple.

if you want a very visual way to transform JSON/XML/CSV/Excel etc in a pipeline it might also be worth looking at Easy Data Transform.

Reply on news.ycombinator.com