app.py

from functools import partial, lru_cache

import duckdb
import gradio as gr
import json
import pandas as pd
import pyarrow as pa
import pyarrow.compute as pc
import requests
from huggingface_hub import HfApi

READ_PARQUET_FUNCTIONS = ("dd.read_parquet", "pd.read_parquet")
EMPTY_TABLE = pa.Table.from_pylist([{str(i): "" for i in range(4)}] * 10)
EMPTY_DF: pd.DataFrame = EMPTY_TABLE.to_pandas()
NUM_ROWS = 20
MAX_NUM_COLUMNS = 20
NUM_TRENDING_DATASETS = 10
NUM_USER_DATASETS = 10
css = """
.transparent-dropdown, .transparent-dropdown .container .wrap, .transparent-accordion  {
    background: var(--body-background-fill);
}
input {
  -webkit-user-select: none;
  -moz-user-select: none;
  -ms-user-select: none;
  user-select: none;          
}
.cell-menu-button {
    z-index: -1;
}
thead {
    display: none;
}
.secondary-wrap:has(input[aria-expanded="true"]) {
    background: var(--table-odd-background-fill);
}
.secondary-wrap:has(input[aria-expanded="true"])::after {
    content: '↵';
    margin-right: var(--size-10);
    border-width: 1px;
    border-color: var(--block-border-color);
    border-radius: .23rem;
    background-color: #141c2e;
    padding-left: 2px;
    font-size: .75rem;
    color: var(--block-title-text-color);
}
.gradio-container {
    padding: var(--size-4) 0 !important;
    max-width: 98% !important;
}
"""
js = """
function load() {
    // Set DataFrame readonly
    MutationObserver = window.MutationObserver || window.WebKitMutationObserver;
    var observer = new MutationObserver(function(mutations, observer) {
        // fired when a mutation occurs
        document.querySelectorAll('.readonly-dataframe div .table-wrap button svelte-virtual-table-viewport table tbody tr td .cell-wrap input').forEach(i => i.setAttribute("readonly", "true"));
    });
    // define what element should be observed by the observer
    // and what types of mutations trigger the callback
    observer.observe(document, {
        subtree: true,
        childList: true
    });

    // Run query on Enter in transform dropdown
    document.querySelectorAll("input").forEach(i => {
        if (i.parentElement.parentElement.parentElement.parentElement.parentElement.classList.contains("transform_dropdown")) {
            i.onkeydown = (event) => {
                if (event.code == "Enter") {
                    document.getElementById("run_button").click();
                }
            }
        }
    })
}
"""
text_functions_df = pd.read_csv("text_functions.tsv", delimiter="\t")
date_functions_df = pd.read_csv("date_functions.tsv", delimiter="\t")
list_functions_df = pd.read_csv("list_functions.tsv", delimiter="\t")
numeric_functions_df = pd.read_csv("numeric_functions.tsv", delimiter="\t")
time_functions_df = pd.read_csv("time_functions.tsv", delimiter="\t")
timestamp_functions_df = pd.read_csv("timestamp_functions.tsv", delimiter="\t")

@lru_cache(maxsize=3)
def duckdb_sql(query: str) -> duckdb.DuckDBPyRelation:
    return duckdb.sql(query)

def prepare_function(func: str, placeholders: list[str], column_name: str) -> str:
    prepared_func = func.split("(", 1)
    for placeholder in placeholders:
        if placeholder in prepared_func[-1]:
            prepared_func[-1] = prepared_func[-1].replace(placeholder, column_name, 1)
            return "(".join(prepared_func)
    else:
        return None

def prettify_df(df: pd.DataFrame):
    return df.apply(lambda s: s.apply(str))

def get_prepared_functions_from_table(table: pa.Table) -> dict[str, list[str]]:
    prepared_functions = {}
    for field in table.schema:
        if pa.types.is_integer(field.type) or pa.types.is_floating(field.type):
            prepared_functions[field.name] = [prepare_function(numeric_func, ["x"], field.name) for numeric_func in numeric_functions_df.Name]
        elif pa.types.is_string(field.type):
            prepared_functions[field.name] = [prepare_function(text_func, ["string"], field.name) for text_func in text_functions_df.Name]
            # try parsing json
            if pc.all(pc.starts_with(table[field.name], "{")).as_py() or pc.all(pc.starts_with(table[field.name], "[")).as_py():
                try:
                    json_parsed_table = pa.Table.from_pylist([{field.name: json.loads(row)} for row in table[field.name].to_pylist()])
                    parsed_type = str(duckdb.from_arrow(json_parsed_table).dtypes[0])
                    prepared_functions[field.name] = [f"CAST({field.name} as {parsed_type})"] + prepared_functions[field.name]
                except Exception:
                    pass
        elif pa.types.is_date(field.type):
            prepared_functions[field.name] = [prepare_function(date_func, ["startdate", "date"], field.name) for date_func in date_functions_df.Name]
        elif pa.types.is_list(field.type):
            prepared_functions[field.name] = [prepare_function(list_func, ["list"], field.name) for list_func in list_functions_df.Name]
        elif pa.types.is_time(field.type):
            prepared_functions[field.name] = [prepare_function(time_func, ["starttime", "time"], field.name) for time_func in time_functions_df.Name]
        elif pa.types.is_timestamp(field.type):
            prepared_functions[field.name] = [prepare_function(timestamp_func, ["startdate", "timestamp"], field.name) for timestamp_func in timestamp_functions_df.Name]
        elif pa.types.is_struct(field.type):
            prepared_functions[field.name] = [f"{field.name}.{subfield.name}" for subfield in field.type.fields]
        else:
            prepared_functions[field.name] = []
        prepared_functions[field.name] = [prepared_function for prepared_function in prepared_functions[field.name] if prepared_function]
    return prepared_functions

with gr.Blocks(css=css, js=js) as demo:
    loading_codes_json = gr.JSON(visible=False)
    dataset_subset_split_textbox = gr.Textbox(visible=False)
    input_table_state = gr.State()
    run_button = gr.Button(visible=False, elem_id="run_button")
    with gr.Row():
        with gr.Column():
            gr.Markdown("# <p style='text-align:center;'>🐥 DuckDB Spreadsheets 📝</p>\n\n<p style='text-align:center;'>Edit any dataset on Hugging Face (full list <a href='https://huggingface.co/datasets' target='_blank'>here</a>) using DuckDB functions (documentation <a href='https://duckdb.org/docs/sql/functions/overview' target='_blank'>here</a>)</p>")
    with gr.Group():
        with gr.Row():
            dataset_dropdown = gr.Dropdown(label="Dataset", allow_custom_value=True, scale=10)
            subset_dropdown = gr.Dropdown(info="Subset", allow_custom_value=True, show_label=False, visible=False, elem_classes="transparent-dropdown")
            split_dropdown = gr.Dropdown(info="Split", allow_custom_value=True, show_label=False, visible=False, elem_classes="transparent-dropdown")
            gr.LoginButton()
        with gr.Row():
            transform_dropdowns = [gr.Dropdown(choices=[column_name] + [prepare_function(text_func, "string", column_name) for text_func in text_functions_df.Name if "string" in text_func], value=column_name, container=False, interactive=True, allow_custom_value=True, visible=True, elem_classes="transform_dropdown") for column_name in EMPTY_DF.columns]
            transform_dropdowns += [gr.Dropdown(choices=[None], value=None, container=False, interactive=True, allow_custom_value=True, visible=False, elem_classes="transform_dropdown") for _ in range(MAX_NUM_COLUMNS - len(transform_dropdowns))]
        dataframe = gr.DataFrame(EMPTY_DF, column_widths=[f"{1/len(EMPTY_DF.columns):.0%}"] * len(EMPTY_DF.columns), interactive=True, elem_classes="readonly-dataframe")
        with gr.Accordion("Show DuckDB SQL command", open=False, elem_classes="transparent-accordion"):
            code_markdown = gr.Markdown()

    def show_subset_dropdown(dataset: str):
        if dataset and "/" not in dataset.strip().strip("/"):
            return []
        resp = requests.get(f"https://datasets-server.huggingface.co/compatible-libraries?dataset={dataset}", timeout=3).json()
        loading_codes = ([lib["loading_codes"] for lib in resp.get("libraries", []) if lib["function"] in READ_PARQUET_FUNCTIONS] or [[]])[0] or []
        subsets = [loading_code["config_name"] for loading_code in loading_codes]
        subset = (subsets or [""])[0]
        return dict(choices=subsets, value=subset, visible=len(subsets) > 1, key=hash(str(loading_codes))), loading_codes

    def show_split_dropdown(subset: str, loading_codes: list[dict]):
        splits = ([list(loading_code["arguments"]["splits"]) for loading_code in loading_codes if loading_code["config_name"] == subset] or [[]])[0]
        split = (splits or [""])[0]
        return dict(choices=splits, value=split, visible=len(splits) > 1, key=hash(str(loading_codes) + subset))

    def show_input_dataframe(dataset: str, subset: str, split: str, loading_codes: list[dict]):
        pattern = ([loading_code["arguments"]["splits"][split] for loading_code in loading_codes if loading_code["config_name"] == subset] or [None])[0]
        if dataset and subset and split and pattern:
            table = duckdb_sql(f"SELECT * FROM 'hf://datasets/{dataset}/{pattern}' LIMIT {NUM_ROWS}").arrow()
        else:
            table = EMPTY_TABLE
        prepared_functions = get_prepared_functions_from_table(table)
        new_transform_dropdowns = [dict(choices=[column_name] + prepared_functions[column_name], value=column_name, container=False, interactive=True, allow_custom_value=True, visible=True) for column_name in table.column_names]
        new_transform_dropdowns += [dict(choices=[None], value=None, container=False, interactive=True, allow_custom_value=True, visible=False) for _ in range(MAX_NUM_COLUMNS - len(new_transform_dropdowns))]
        df = table.to_pandas()
        return [table, dict(value=prettify_df(df), column_widths=[f"{1/len(df.columns):.0%}"] * len(df.columns))] + new_transform_dropdowns
    
    def set_dataframe(dataset: str, subset: str, split: str, loading_codes: list[dict], input_table: pa.Table, df: pd.DataFrame, *transforms, show_warning=True):
        try:
            table = duckdb.sql(f"SELECT {', '.join(transform for transform in transforms if transform)} FROM input_table;").arrow()
        except Exception as e:
            if show_warning:
                gr.Warning(f"{type(e).__name__}: {e}")
            return {
                dataframe: df
            }
        prepared_functions = get_prepared_functions_from_table(table)
        new_transform_dropdowns = [dict(choices=list({original_column_name: None, column_name: None}) + prepared_functions[column_name], value=column_name, container=False, interactive=True, allow_custom_value=True, visible=True) for original_column_name, column_name in zip(input_table.column_names, table.column_names)]
        new_transform_dropdowns += [dict(choices=[None], value=None, container=False, interactive=True, allow_custom_value=True, visible=False) for _ in range(MAX_NUM_COLUMNS - len(new_transform_dropdowns))]
        pattern = ([loading_code["arguments"]["splits"][split] for loading_code in loading_codes if loading_code["config_name"] == subset] or [None])[0]
        return {
            dataframe: prettify_df(table.to_pandas()),
            **dict(zip(transform_dropdowns, [gr.Dropdown(**new_transform_dropdown) for new_transform_dropdown in new_transform_dropdowns])),
            code_markdown: (
                "```sql\n"
                + f"SELECT {', '.join(new_transform_dropdown['value'] for new_transform_dropdown in new_transform_dropdowns if new_transform_dropdown['value'])} "
                + f"FROM 'hf://datasets/{dataset}/{pattern}' LIMIT {NUM_ROWS};"
                + "\n```"
            ) if pattern else "",
        }

    for column_index, transform_dropdown in enumerate(transform_dropdowns):
        transform_dropdown.select(partial(set_dataframe, show_warning=False), inputs=[dataset_dropdown, subset_dropdown, split_dropdown, loading_codes_json, input_table_state, dataframe] + transform_dropdowns, outputs=[dataframe, code_markdown] + transform_dropdowns)

    run_button.click(set_dataframe, inputs=[dataset_dropdown, subset_dropdown, split_dropdown, loading_codes_json, input_table_state, dataframe] + transform_dropdowns, outputs=[dataframe, code_markdown] + transform_dropdowns)

    @demo.load(outputs=[dataset_dropdown, loading_codes_json, subset_dropdown, split_dropdown, input_table_state, dataframe, code_markdown] + transform_dropdowns)
    def _fetch_datasets(request: gr.Request, oauth_token: gr.OAuthToken | None):
        api = HfApi(token=oauth_token.token if oauth_token else None)
        datasets = list(api.list_datasets(limit=NUM_TRENDING_DATASETS, sort="trendingScore", direction=-1, filter=["format:parquet"]))
        if oauth_token and (user := api.whoami().get("name")):
            datasets += list(api.list_datasets(limit=NUM_USER_DATASETS, sort="trendingScore", direction=-1, filter=["format:parquet"], author=user))
        dataset = request.query_params.get("dataset") or datasets[0].id
        subsets, loading_codes = show_subset_dropdown(dataset)
        splits = show_split_dropdown(subsets["value"], loading_codes)
        input_table, input_dataframe, *new_transform_dropdowns = show_input_dataframe(dataset, subsets["value"], splits["value"], loading_codes)
        pattern = ([loading_code["arguments"]["splits"][splits["value"]] for loading_code in loading_codes if loading_code["config_name"] == subsets["value"]] or [None])[0]
        return {
            dataset_dropdown: gr.Dropdown(choices=[dataset.id for dataset in datasets], value=dataset),
            loading_codes_json: loading_codes,
            subset_dropdown: gr.Dropdown(**subsets),
            split_dropdown: gr.Dropdown(**splits),
            input_table_state: input_table,
            dataframe: gr.DataFrame(**input_dataframe),
            **dict(zip(transform_dropdowns, [gr.Dropdown(**new_transform_dropdown) for new_transform_dropdown in new_transform_dropdowns])),
            code_markdown: (
                "```sql\n"
                + f"SELECT {', '.join(new_transform_dropdown['value'] for new_transform_dropdown in new_transform_dropdowns if new_transform_dropdown['value'])} "
                + f"FROM 'hf://datasets/{dataset}/{pattern}' LIMIT {NUM_ROWS};"
                + "\n```"
            ) if pattern else "",
        }
    
    @dataset_dropdown.select(inputs=dataset_dropdown, outputs=[loading_codes_json, subset_dropdown, split_dropdown, input_table_state, dataframe, code_markdown] + transform_dropdowns)
    def _show_subset_dropdown(dataset: str):
        subsets, loading_codes = show_subset_dropdown(dataset)
        splits = show_split_dropdown(subsets["value"], loading_codes)
        input_table, input_dataframe, *new_transform_dropdowns = show_input_dataframe(dataset, subsets["value"], splits["value"], loading_codes)
        pattern = ([loading_code["arguments"]["splits"][splits["value"]] for loading_code in loading_codes if loading_code["config_name"] == subsets["value"]] or [None])[0]
        return {
            loading_codes_json: loading_codes,
            subset_dropdown: gr.Dropdown(**subsets),
            split_dropdown: gr.Dropdown(**splits),
            input_table_state: input_table,
            dataframe: gr.DataFrame(**input_dataframe),
            **dict(zip(transform_dropdowns, [gr.Dropdown(**new_transform_dropdown) for new_transform_dropdown in new_transform_dropdowns])),
            code_markdown: (
                "```sql\n"
                + f"SELECT {', '.join(new_transform_dropdown['value'] for new_transform_dropdown in new_transform_dropdowns if new_transform_dropdown['value'])} "
                + f"FROM 'hf://datasets/{dataset}/{pattern}' LIMIT {NUM_ROWS};"
                + "\n```"
            ) if pattern else "",
        }
    
    @subset_dropdown.select(inputs=[dataset_dropdown, subset_dropdown, loading_codes_json], outputs=[split_dropdown, input_table_state, dataframe, code_markdown] + transform_dropdowns)
    def _show_split_dropdown(dataset: str, subset: str, loading_codes: list[dict]):
        splits = show_split_dropdown(subset, loading_codes)
        input_table, input_dataframe, *new_transform_dropdowns = show_input_dataframe(dataset, subset, splits["value"], loading_codes)
        pattern = ([loading_code["arguments"]["splits"][splits["value"]] for loading_code in loading_codes if loading_code["config_name"] == subset] or [None])[0]
        return {
            split_dropdown: gr.Dropdown(**splits),
            input_table_state: input_table,
            dataframe: gr.DataFrame(**input_dataframe),
            **dict(zip(transform_dropdowns, [gr.Dropdown(**new_transform_dropdown) for new_transform_dropdown in new_transform_dropdowns])),
            code_markdown: (
                "```sql\n"
                + f"SELECT {', '.join(new_transform_dropdown['value'] for new_transform_dropdown in new_transform_dropdowns if new_transform_dropdown['value'])} "
                + f"FROM 'hf://datasets/{dataset}/{pattern}' LIMIT {NUM_ROWS};"
                + "\n```"
            ) if pattern else "",
        }
    
    @split_dropdown.select(inputs=[dataset_dropdown, subset_dropdown, split_dropdown, loading_codes_json], outputs=[input_table_state, dataframe, code_markdown] + transform_dropdowns)
    def _show_input_dataframe(dataset: str, subset: str, split: str, loading_codes: list[dict]) -> pd.DataFrame:
        input_table, input_dataframe, *new_transform_dropdowns = show_input_dataframe(dataset, subset, split, loading_codes)
        pattern = ([loading_code["arguments"]["splits"][split] for loading_code in loading_codes if loading_code["config_name"] == subset] or [None])[0]
        return {
            input_table_state: input_table,
            dataframe: gr.DataFrame(**input_dataframe),
            **dict(zip(transform_dropdowns, [gr.Dropdown(**new_transform_dropdown) for new_transform_dropdown in new_transform_dropdowns])),
            code_markdown: (
                "```sql\n"
                + f"SELECT {', '.join(new_transform_dropdown['value'] for new_transform_dropdown in new_transform_dropdowns if new_transform_dropdown['value'])} "
                + f"FROM 'hf://datasets/{dataset}/{pattern}' LIMIT {NUM_ROWS};"
                + "\n```"
            ) if pattern else "",
        }


if __name__ == "__main__":
    demo.launch()