from Hacker News

Kurtosis: A composable build system for reproducible environments

by dloss on 10/28/23, 8:22 AM with 30 comments

• by hobofan on 10/28/23, 7:39 PM

  Just to spare anyone the time: When they say "reproducible" they mean "can be done again" and not "can be done again with a consistent result".

  No lockfiles, no checksums, URLs that may disappear and change at any time strewn around everywhere. There is none of the many nowadays common mechanisms in place that would ensure reproducibility (as the term is used in other build systems like Bazel or Nix).

• by wwwpatdelcom on 10/28/23, 2:12 PM

  As a both devops and ML professional at my job (it's a small company), I viscerally cringe at the same kurtosis for this company (which I just learned about). There could be some kind of latent jealously/projection causing that cringe, but here's my rationale (could be very specific to my career):

  * Devops folks don't seem to tend to like math and often got there by practicing, "computers, IT, having fun hooking things together and getting them running."

  * Data science folks don't tend to like devops and prefer to bash around on a jupyter notebook that's already given to them and then maybe extract that python and see if it runs, but they tend to come from more of a science background and got into python as a hobby or incidentally. They do not like bashing around and getting things running.

  So now this company is combining a math term that has a specific meaning to an ML ops space, which is going to cause confusion.

  Different sets of data can have different kurtosis measures. Sets can be platykurtic (flat gaussian curve or high kurtosis) or leptokurtic (tall gaussian curve, low kurtosis).

  Now this company is coming in and telling a bunch of devops people, "Kurtosis means helm but automatically migrate data too." So they are applying the idea of, "leptokurtic deployments," presumably with the metric being, variation between the code and data parameters on those servers. Data science people who are told about it from devops people are going to initially hear, "somehow dealing with cleaning the data, like an ETL pipeline, perhaps an Airflow with data cleaning tools built in or something."

  It's very confusing and not helpful to customers, I hate it. There are going to be meetings where ML/Devops people are very confused.

  Naming is hard though -- but I wish they would have gone with something like, "platypus" and just have a cute little platypus baby as the logo and say, "yeah we liked the word platykurtic because we like making things regular and platykurtic sounds like platypus."

• by SrslyJosh on 10/28/23, 5:37 PM

  That "Web3" link at the top of the page is a big time-saver.

• by gavinhoward on 10/28/23, 8:30 PM

  I'm making a build system too.

  Every time I see one announced, I check a few things.

  The first thing I check is the build language. Usually, that indicates to me whether or not the build system would target the same audience that I eventually will.

  In this case, they chose Starlark.

  Whew! We are not going to target the same audience.

• by FredPret on 10/28/23, 6:41 PM

  I'm naming my next startup "percentile"

• by dpflan on 10/28/23, 1:14 PM

  "A composable build system for reproducible environments"

  Is there another meaning of kurtosis that is applicable to this product? Kurtosis is used to describe aspects of probability distributions' tails/flatness.

  Is this a pun on "distributed" systems?

• by remram on 10/28/23, 5:00 PM

  The "QuickStart" link on this page doesn't work for me, I am guessing this is where it should go: https://docs.kurtosis.com/quickstart

• by djupblue on 10/28/23, 11:14 PM

  Is there an example demonstrating what it looks like and how it works?