Altair graphs in a random fashion

[ale-dg]

What happened?

Hi all,

I have been creating several times some graphs using altair with the following code:

Code block 1

alt.Chart(
    df.group_by("season")
    .agg(pl.count("season").alias("count"))
).mark_bar().encode(
    alt.X(
        "season:N",
        title="Season",
        sort=['Spring', 'Summer', 'Autumn', 'Winter'],
    ),
    alt.Y("count:Q", title="Frequency"),
    alt.Color("season:N", title="Season", legend=None),
    [
        alt.Tooltip("count:Q", title="Frequency", format=".3s"),
        alt.Tooltip("season:N", title="Season"),
    ],
).properties(
    width=600, height=300, title="Number of trips per season"
)

Every time I run it I am getting different results, not in the amounts but in the order the graph is rendered, even though I have set an specific one in the alt.X options, see an example below:

Image

I have found a workaround it, also aggregating some data from Polars, although I am curious about the behaviour of altair, as I am not sure why this is happening. I leave you the code I am using for correctly rendering the graph:

Code block 2

alt.Chart(
    df.group_by("season")
    .agg(pl.count("season").alias("count"))
    .with_columns(
        pl.when(pl.col("season") == "Spring")
        .then(1)
        .when(pl.col("season") == "Summer")
        .then(2)
        .when(pl.col("season") == "Autumn")
        .then(3)
        .otherwise(4)
        .alias("season_order")
    )
).mark_bar().encode(
    alt.X(
        "season:N",
        title="Season",
        sort=alt.EncodingSortField(field="season_order", order="ascending"),
    ),
    alt.Y("count:Q", title="Frequency"),
    alt.Color("season:N", title="Season", legend=None),
    [
        alt.Tooltip("count:Q", title="Frequency", format=".3s"),
        alt.Tooltip("season:N", title="Season"),
    ],
).properties(
    width=600, height=300, title="Number of trips per season"
)

Thank you for your help!

Best

What would you like to happen instead?

It should be rendering in the sort order specified, not in a random fashion

Which version of Altair are you using?

5.5.0

[dangotbanned]

@ale-dg can you provide the source df for this repro please?

If you omitted it because you are unable to share it, could you instead rewrite your examples with dummy data that also demonstrates the issue?

[ale-dg]

@dangotbanned thank you for your reply. I didn't attach the source since it is a 1.7 GB parquet file... so really not feasible to attach it.

I did the following code with the Seattle Weather dataset, also using Polars, although this is working fine:

Code block 3

import altair as alt
import polars as pl
from vega_datasets import data

pl.Config.set_tbl_cols(100)
alt.theme.enable("googlecharts")

source = data.seattle_weather()

df = pl.DataFrame(source)

df = df.with_columns(pl.col("date").dt.month().alias("month"))

df = df.with_columns(
    pl.when(pl.col("month").is_between(3, 5))
    .then(pl.lit("Spring"))
    .when(pl.col("month").is_between(6, 8))
    .then(pl.lit("Summer"))
    .when(pl.col("month").is_between(9, 11))
    .then(pl.lit("Autumn"))
    .otherwise(pl.lit("Winter"))
    .alias("season")
)

alt.Chart(
    df.group_by("season").agg(pl.count("season").alias("count"))
).mark_bar().encode(
    alt.X("season:N", title="Season", sort=["Spring", "Summer", "Autumn", "Winter"]),
    alt.Y("count", title="Frequency"),
    alt.Color("season:N", legend=None),
    [
        alt.Tooltip("season:N", title="Season"),
        alt.Tooltip("count", title="Frequency", format=".3s"),
    ],
).properties(
    title="Count of records per season", width=900, height=300
)

Look below for some screenshots of the original issue. I copied and paste the code for generating the graph in different Jupyter cells and the issue happens everytime. I should also mention it is not only happening with the "season" column, it is also happening with any other column where I try to be specific on the order the graph should appear.

Hope this helps! Otherwise, please let me know.

Best

4 Images

[dangotbanned]

@ale-dg does your original preserve the order if you change the encoding type?

Original code block

#3714 (comment)

alt.Chart(
    df.group_by("season")
    .agg(pl.count("season").alias("count"))
).mark_bar().encode(
    alt.X(
        "season:N",
        title="Season",
        sort=['Spring', 'Summer', 'Autumn', 'Winter'],
    ),
    alt.Y("count:Q", title="Frequency"),
    alt.Color("season:N", title="Season", legend=None),
    [
        alt.Tooltip("count:Q", title="Frequency", format=".3s"),
        alt.Tooltip("season:N", title="Season"),
    ],
).properties(
    width=600, height=300, title="Number of trips per season"
)

alt.X(
        # "season:N",
        "season:O",
        title="Season",
        sort=['Spring', 'Summer', 'Autumn', 'Winter'],
    )

Per encoding-data-types:

Data Type	Shorthand Code	Description
quantitative	Q	a continuous real-valued quantity
ordinal	O	a discrete ordered quantity
nominal	N	a discrete unordered category
temporal	T	a time or date value
geojson	G	a geographic shape

To me, it sounds like you know that "season" has an order.
But that is at odds with telling altair that it is an unordered category.

---

Alternatives

My other suggestions on the polars-side if the above doesn't fix it:

Tell polars to preserve order (pl.DataFrame.group_by)

# df.group_by("season").agg(pl.count("season").alias("count"))
df.group_by("season", maintain_order=True).agg(pl.count("season").alias("count"))

Using GroupBy.len instead of pl.count

# df.group_by("season").agg(pl.count("season").alias("count"))
df.group_by("season").len("count")

---

Hope something here can help!