Starlark Transformations

Qri (“query”) is about datasets. A transformation is a repeatable script for generating a dataset. Starlark. This package implements starlark as a transformation syntax. Starlark transformations are about as close as one can get to the full power of a programming language as a transformation syntax. Often you need this degree of control to generate a dataset.

Typical examples of a starlark transformation include:

  • combining paginated calls to an API into a single dataset
  • downloading unstructured structured data from the internet to extract

So let’s dive in and learn about transforms!

Table of Contents

1. Transformation Basics

2. Transforms that Download Data

1.1 What’s a Transformation?

Before we get started, let’s understand what we mean by “transform”. Here’s the technical definition of a transform in Qri:

A Transformation is a forward transition from one snapshot to another snapshot.

In plain English: transformations are how datasets change. There are two types of transformation: manual and scripted. Manual transforms are direct manipulations of data, scripted transformations use code to make changes.

There are some rules to how transformations work:

  • transforms must mutate one or more non-computed fields of a dataset
  • only one type of transform can be applied to any field per transform
  • transforms can use one or more types of mutations to determine the next snapshot

It’s totally ok if that sounds like nonsense for now. We’ll be walking through all of this with real examples in this tutorial.

1.2 Manual Transforms

Before getting into scripts, let’s create a dataset using only manual transformations. Manual transforms work by providing values directly to Qri. Let’s start by manually creating a dataset. First we’ll create a new (very simple) json file: an array of rational numbers called body.json:


Next in the same folder we’ll create a new file called dataset.yaml with the following contents:

name: rational_numbers
  title: rational number series
bodypath: body.json

From a terminal, navigate to the folder that contains that file, and save it to Qri with qri save.

$ cd path/to/that_folder
$ qri save --file=dataset.yaml
saved dataset b5/rational_numbers

Ok cool, you’ve just created a new dataset with a manual transformation that lists rational numbers. It’s “manual” because you provided values directly to Qri. The “transform part” transitioned from nothing (an empty dataset) to a an initial snapshot of a datset.

1.3 Transform Scripts

We can think of the above manual transform as a series of assignments in a single function call. Written out as code, above example is telling Qri to do the following:

def human_transform(ds):
  ds.set_meta("title", "rational number series")

When you save a dataset, qri calls up the previous version (or crates an empty dataset when there is no previous version), and then applies all the changes provided to get to a new version.

There are many situations where manual transforms are the right option (or the only option) for changing a dataset. But we’re using a computer, and if we can describe the changes we want to make as code, the computer can do more work for us. That’s where transform scripts come in.

With a transform script, instead of making manual changes, an algorithm automates changes to fields of a dataset with programmatic instructions.

Before we get into what’s going on, let’s actually try this out.

Transform scripts are written in starlark. Starlark is a subset of python, so if you can write python, you can write starlark. If you can’t write python (or starlark), that’s ok! We’ll circle back later & explain the transform function, but for now let’s just copy-paste our way to victory. From the same folder, create a new file called and save this into it:

def transform(ds, ctx):
  ds.set_meta("title", "rational number series")
  # use the range function to automate setting the dataset body
  # range(1,11) will produce an array of numbers: [1,2,3,4,5,6,7,9,10]
  # which is the same as our manual transform, but with less typing!

note: if you’re using a text editor and want syntax highlighting (colored text), try setting your editor to ‘python’ syntax.

This is a script that does the exact same thing as our manual transform from earlier. To use the script, let’s modify our dataset.yaml use the script, deleting the meta and body components, and adding a new transform component that specifies our script file. Once dataset.yaml looks like this, save the file:

name: rational_numbers
  scriptpath: transform.yaml

You can delete stuff in this file, because it’s stored in Qri!

Now let’s save a new snapshot to qri:

$ qri save --file dataset.yaml
error saving: no changes detected

Wait, we got an error. what gives? This is because the result of running the transform didn’t change the dataset. Transforms have to describe changes. This is a super important feature of Qri, and transforms. If nothing changes, Qri can tell you as much, and avoid creating unnecessary versions of a dataset.

To get this to work, let’s change something! Let’s open up our file and write a script that adds more numbers to our body:

def transform(ds, ctx):
  ds.set_meta("title", "rational number series")
  # this time set the body to 1-1000 instead:

And re-run save:

$ qri save --file dataset.yaml
dataset saved: b6/[email protected]

Congrats! you’ve just run a transform with a dataset of the numbers 1 to 1000, without having to type the numbers by hand.

The transform function

Ok, we can’t avoid the issue any longer, time to understand what’s going on in this script:

def transform(ds, ctx):
  ds.set_meta("title", "rational number series")
  # this time set the body to 1-1000 instead:

This script defines a special function: transform that Qri knows to look for. There are others, but transform is the main one (literally, if you’re coming from an engineering background, transform is Qri’s main). The transform function accepts two arguments: ds and ctx: * ds represents the current dataset snapshot. Your mission, should you choose to accept it, is to change ds in some way * ctx represents the transform context, it keeps info used while the transform script is running.

When we run qri save with a transform script specified, Qri will load it up and look for the transform function, which must be defined. When Qri finds that function, it loads up the existing version of the dataset (or makes an empty dataset if there’s no history), and passes it to ds. Whatever changes are made to ds via method calls like set_body and set_meta are applied to the dataset, and the result is committed as a new snapshot.

Its possible to mix manual transforms and scripted transforms, but they can’t affect the same parts of a dataset. We’ll cover mixing transform types in another tutorial.

2. Transforms that Download Data

Transforms have another special function called download that lets you create super-powered datasets that draw from the world wide web. download can do all sorts of stuff like grab resources from APIs, fetch & parse HTML, or pull raw csv data off the web. Combined with Qri versioning, you can make a dataset that knows how to update itself, and only records updates when the external resource changes.

2.1 Download function

In this section we are going to talk about the Qri function download. The download function is a special function that gives your script access to the internet. The download function is the only place where you have access to the web.

The download function is always run before the transform step, and places it’s results in the transform context. The dataset returned from the download function, gets passed as the dataset parameter in the transform function.

2.2 Config and Secrets

When we came up with the idea of including transforms in Qri, the thing we were most excited about was a transforms ability to be customized for the person running it. In order to have customizability, we needed a way to configure a transform script. For example, if there is a dataset that has a call to the github api, that can pull down the stats from one of my projects, but I want to also use that transform to pull down stats from a second project, one of the variables in my transform would probably be repo_name.

Related to configuration are secrets. Often, when we try to get information from an api, that api requires us to have a special key that is only associated with our identity. It’s often private and should not be shared in the dataset itself or made public in any way.

This is where the config and secrets comes in. config and secrets are both part of the transform dataset component.

To illustrate, we’ll build an example that grabes the last 100 League of Legends matches a specific player (in this case called summoner) has played in a specific region (in this case, North America). Let’s create a new folder called lol_last_100_matches, and a new dataset.yaml file within that folder:

# lol_last_100_matches/dataset.yaml
name: league_player_matches
  title: Dataset created using the qri starlark tutorial. Pings the Riot Games (creator of the computer game League of Legends) api, gets a summoner's account id, and then a list of their last 100 matches.
    summoner: sørenbjerg
    region: na1

Config is set right on the transform component. Secrets, on the other hand, should be provided when calling the command:

$ qri save --file=dataset.yaml --secrets=api_key,*******************

Note, if you want to run this transform yourself, you will need to head over to the Riot Games developer website, and create a login. Then, you can generate your own api key right from your developer dashboard. You will replace the series of ***** in the command with your own api key.

Okay, so that’s how you add a config variable and a secrets variable into the transform, but how do you actually use it in the transform file? That’s where transform context comes in.

2.3 Modules and qri

Chances are, if you are trying to do something cool with Qri, you will need more than just the basic functionality we’ve shown you so far.

You can also import modules from the starlark standard library (Starlib we have been working on. Here is our reference page that details each module and each function within that module.

For now, let’s look at the qri module. Here is how you load a module into a transform:

# lol_last_100_matches/
load("", "qri")

def download(ctx):
  return ds

Great, now that we’ve loaded the qri module, let’s actually use it to get the summoner name, region, and api_key, from the dataset.

def download(ctx):
  summoner = ctx.get_config("summoner")
  region = ctx.get_config("region")
  api_key = ctx.get_secret("api_key")
  print(summoner) # prints "sørenbjerg"
  print(region) # prints "na1"
  print(api_key) # prints the api_key

def transform(ds,ctx):
  # nothing yet

Head over to the terminal. Change directories until you are in your lol_last_100_matches folder. We are going to use the --dry-run flag. This will allow us to see the output of the transform, without actually saving it to our Qri node.

$ qri save --file dataset.yaml me/lol_last_100_matches --secrets=api_key,******** --dry-run

2.4 http module

Now that we can get config and secrets variables, let’s use those to grab some data from an API endpoint using the http package.

The http package can only be used in the download function. You do not have access to the network in any the transform function. If you try to download something in the transform function, you will get an error.

You have access to the get, put, post, delete, patch, and options methods from the http module. We are going to use the get method to grab some json for the Riot Games json api.

Note: you can use the text, content, and json methods on a response to get the response body. text and content will return a string representation, json will convert it to json. Please see the starlib reference page for more info.

load("", "qri")
load("", "http")

def download(ctx):
  # get config and secrets variables from dataset
  summoner = ctx.get_config("summoner")
  region = ctx.get_config("region")
  api_key = ctx.get_secret("api_key")
  # get response from api endpoint
  res = http.get("https://" + region + "" + summoner + "?api_key=" + api_key)
  # get json from body, get accountId from json
  # note: when the response gets converted to json, it automatically gives the 
  # accountId and id fields a float type. If we were to convert straight to 
  # a string, it would appear in scientific notation. To combat this, I am 
  # casting it to int type, before finally settling on string type.
  data = res.json()
  accountId = str(int(data["accountId"]))
  # get match data and set as the body of the dataset
  res = http.get("https://" + region + "" + accountId + "?api_key=" + api_key)
  # the response is a python dictionary with the dictionary key "matches"
  # that points to a list of dictionaries. Each item in the list contains 
  # information on a specific match. We only want to keep the actual list
  # of matches:
  matches = res.json()["matches"]
  print(matches) # prints a long list of dictionaries containing match data
  return matches

def transform(ds, ctx):
  # get matches from context
  matches =
  # set the body to our matches data

First double check that this works by running

$ qri save --file dataset.yaml --secrets=api_key,****** --dry-run
🏃🏽‍♀️ dry run
🤖 executing transform
📡 running download...
⚙️  running transform...
# HERE THE OUTPUT OF THE `print(matches)` CALL
✅ transform complete
created new dataset me/[email protected]TXF6LzpCFK87Ykq7WR7hjzCvNXWGXZ2ssJZwRiPiMSN9/map/QmV9jSavMgEFnC2METNKUvJLb4sVvoMTf6EthuAKmqzbfg

If you are running into problems, double check that your api key is up to date!

To get this dataset into your qri node, run the same command, without the --dry-run flag.

2.4 html module

Let’s take a cursory look at the html module. The html module allows you use methods to grab elements from an html page, much like you would using jquery. Take a look at the starlib reference page to find out more.

Let’s go to wikipedia, and get a list of all the languages that you can read wikipedia in!

We will download the main wikipedia page, parse it using the html method, then navigate down to the element, get the language from the ‘title’ attribute, and add it to the list of languages.

load("", "html")
load("", "http")

def download(ctx):
  res = http.get("")
  doc = html(res.content())
  # from inspecting the contents of the wikipedia page, I was able to 
  # determine that the list of languages was located in a div who's id 
  # is `#p-lang` 
  langElems = doc.find("#p-lang").find("li")
  langs = []
  for i in range(0, langElems.len()):
    # get the ith element in the list of <li>'s
    li = langElems.eq(i)
    # list of the children of the <li>, in this case it is a list of one
    # element, an <a> tag
    alist = li.children()
    # There is only one <a> in the list, but it is still a list
    # to get that first <a> element:
    a = alist.first()
    # the "title" attribute in the <a> element contains the language, written
    # in english
    language = a.attr("title")
    # append this to the list of languages
    # this can all be done, alittle more confusingly, in one line:
    # langs.append(langElems.eq(i).children().first().attr("title"))
  return langs

def transform(ds, ctx):
  langs =

To learn more about our starlark standard library, check out the reference page which details each module and all of it’s methods.