Quick start

This example reads laboratory observations, removes rows without an OD600 measurement, summarizes the remaining values by sample, and publishes a new Arrow table.

1. Declare the block ports

Create a Python processing block with these ports:

Direction	Name	Type
Input	rows	Arrow table
Output	summary	Arrow table

Port names are part of the block's contract. Use letters, numbers, and underscores; do not use spaces or hyphens.

2. Add the processing script

import pyarrow.compute as pc

import flasktrack as ft

rows = ft.input_table("rows")

required = {"sample_id", "od600"}
missing = required.difference(rows.column_names)
if missing:
    raise ValueError(f"Missing required columns: {sorted(missing)}")

clean = rows.filter(pc.is_valid(rows["od600"]))
summary = clean.group_by("sample_id").aggregate(
    [
        ("od600", "mean"),
        ("od600", "stddev"),
        ("od600", "count"),
    ]
)

ft.output_table("summary", summary)
ft.log(f"Published {summary.num_rows} sample summaries")

3. Connect the block

Connect an Arrow table output from a source or earlier processing block to rows. Connect summary to a compatible downstream block or output.

The connection type must match at both ends. An Arrow table cannot be connected directly to a JSON or file port.

4. Test with representative data

Before publishing the pipeline version, test with data that includes:

• several sample identifiers;
• at least one missing OD600 value;
• repeated observations for a sample; and
• the same column names and types expected in production.

Confirm that the output row count, column names, calculated values, and missing value behavior are correct.

5. Review before publishing

Check that the script:

• reads every declared input;
• publishes every declared output;
• uses literal port names exactly as declared;
• does not assume columns that are absent from the input schema; and
• raises a useful error instead of publishing misleading empty output.

For different dataframe libraries, see Tables and dataframes.