Update Chart.transform_filter() to mirror alt.when()

[dangotbanned]

What is your suggestion?

The goal is simply to provide equivalent predicate parsing functionality, that is found in alt.when().

For example, this would be easy to port over:

verbose_composition = (
    (alt.datum.Name == "Name_1")
    & (alt.datum.Color == "Green")
    & (alt.datum.Age == 25)
    & (alt.datum.StartDate == "2000-10-01")
)
when_verbose = alt.when(verbose_composition)
when_concise = alt.when(Name="Name_1", Color="Green", Age=25, StartDate="2000-10-01")

I was reading through transform/filter and looking at the implementation:

.transform_filter() implementation

altair/altair/vegalite/v5/api.py

Lines 2917 to 2924 in 5a6f70d

	if isinstance(filter, Parameter):
	new_filter: dict[str, Any] = {"param": filter.name}
	if "empty" in kwargs:
	new_filter["empty"] = kwargs.pop("empty")
	elif isinstance(filter.empty, bool):
	new_filter["empty"] = filter.empty
	filter = new_filter
	return self._add_transform(core.FilterTransform(filter=filter, **kwargs))

AFAIK, the signature can be changed without impacting any existing code:

altair/altair/vegalite/v5/api.py

Lines 2888 to 2898 in 5a6f70d

	def transform_filter(
	self,
	filter: str
	| Expr
	| Expression
	| Predicate
	| Parameter
	| PredicateComposition
	| dict[str, Predicate | str | list | bool],
	**kwargs: Any,
	) -> Self:

• There are no checks on filter being in **kwargs
• The only "valid" keyword is empty
  • condition() has already been safely updated in this way

I definitely want to do some more testing, but I can't see much that would be blocking closing the gap between the two:

altair/altair/vegalite/v5/api.py

Lines 1214 to 1219 in 5a6f70d

	def when(
	predicate: Optional[_PredicateType] = Undefined,
	*more_predicates: _ComposablePredicateType,
	empty: Optional[bool] = Undefined,
	**constraints: _FieldEqualType,
	) -> When:

Side note

#695 is referenced on that page, but I think it was resolved a while ago?

Also there are references to selection() which was deprecated in 5.0.0

Have you considered any alternative solutions?

• Leave .transform_filter() unchanged
• Accept alt.When (class) as filter
  • Would work by unwrapping When._condition
  • I'd prefer to add the expressiveness, without annotating .transform_filter with When
    • Seems less intuitive to me

[mattijn]

I think I support this request, but to be sure with an example:

• current (first example on https://altair-viz.github.io/user_guide/transform/filter.html

import altair as alt
from altair import datum

from vega_datasets import data
pop = data.population.url

alt.Chart(pop).mark_area().encode(
    x='age:O',
    y='people:Q',
).transform_filter(
    (datum.year == 2000) & (datum.sex == 1)
)

• with this feature request

import altair as alt
from vega_datasets import data

source = data.population.url

alt.Chart(source).mark_area().encode(
    x='age:O',
    y='people:Q',
).transform_filter(
    alt.when(year=2000, sex=1)
)

[dangotbanned]

I think I support this request, but to be sure with an example:

Thanks for the quick response @mattijn

Apologies, I wrote this up yesterday in a bit of a rush.

To clarify, below I've rewritten your comment with what I intended:

Current (first example on transform/filter)

import altair as alt
from altair import datum
from vega_datasets import data

pop = data.population.url

alt.Chart(pop).mark_area().encode(
    x='age:O',
    y='people:Q',
).transform_filter(
    (datum.year == 2000) & (datum.sex == 1)
)

With this feature request

import altair as alt
from vega_datasets import data

source = data.population.url

alt.Chart(source).mark_area().encode(
    x='age:O',
    y='people:Q',
).transform_filter(year=2000, sex=1)

Alternative (2)

Your example corresponds to what I mentioned at the end of the description

Have you considered any alternative solutions?

• Leave .transform_filter() unchanged
• Accept alt.When (class) as filter
  • Would work by unwrapping When._condition
  • I'd prefer to add the expressiveness, without annotating .transform_filter with When
    • Seems less intuitive to me

import altair as alt
from vega_datasets import data

source = data.population.url

alt.Chart(source).mark_area().encode(
    x='age:O',
    y='people:Q',
).transform_filter(
    alt.when(year=2000, sex=1)
)

---

Expanding on why I prefer not to include When

• Using *args, **kwds would correspond directly to polars.DataFrame.filter
  • I feel there is a nice alignment then between:
    • pl.when -> alt.when
    • pl.DataFrame.filter -> alt.Chart.transform_filter
  • This makes it intuitive for users introduced to altair via polars
    • And in the future, maybe the reverse would also be true
• We would not need to add &, |, ~ operators to When
  • E.g. this isn't valid
    • alt.when(year=2000, sex=1) | datum.name == "john"
• Syntactically, there are fewer concepts happening in the first - but it is not less expressive
  • filter-predicates-constraints
  • filter-when-predicates-constraints

[mattijn]

Ah, thanks for clarifying. Sounds like a good idea 👍

[jonmmease]

I'm in favor, we're not using the kwargs for anything currently, and using them for direct equality makes sense.

[dangotbanned]

Experimenting with this now, will publish a branch soon tracking in #3664.

Early findings

• Entirely removing **kwargs breaks no tests/examples
  • Provided empty is added as named keyword-only argument
• Ranged Dot Plot is the only example which uses filter as a keyword
  • Not a deal-breaker
  • Will probably want to issue a deprecation warning and suggest passing as positional
    • Popping filter from **constraints
  • During the deprecation period, we'll have one edge case that cannot support the new syntax
    • .transform_filter(filter=value), will still need to use
    • .transform_filter(alt.datum["filter"] == value)