Apache Arrow — Columnar Data Format

Overview

Apache Arrow, the cross-language columnar memory format for analytics workloads. Helps developers use Arrow for high-performance data interchange between systems, zero-copy reads, and efficient columnar processing in Python (PyArrow) and JavaScript (Arrow JS).

Instructions

PyArrow — Python Interface

python
# src/data/arrow_ops.py — High-performance data operations with PyArrow
import pyarrow as pa
import pyarrow.parquet as pq
import pyarrow.compute as pc
import pyarrow.csv as pcsv

# Create Arrow tables from Python data
table = pa.table({
    "user_id": pa.array([1, 2, 3, 4, 5], type=pa.int64()),
    "name": pa.array(["Alice", "Bob", "Charlie", "Diana", "Eve"]),
    "revenue": pa.array([150.0, 320.5, 89.0, 1200.0, 45.5], type=pa.float64()),
    "signup_date": pa.array([
        "2026-01-15", "2026-01-20", "2026-02-01", "2026-02-10", "2026-03-01"
    ]).cast(pa.date32()),
    "is_active": pa.array([True, True, False, True, False]),
})

# Compute operations (vectorized, no Python loops)
high_value = pc.filter(table, pc.greater(table["revenue"], 100))
total_revenue = pc.sum(table["revenue"]).as_py()    # 1805.0
avg_revenue = pc.mean(table["revenue"]).as_py()     # 361.0
sorted_table = pc.sort_indices(table, sort_keys=[("revenue", "descending")])

# Read/write Parquet files (the standard format for Arrow data)
pq.write_table(table, "users.parquet", compression="zstd")
loaded = pq.read_table("users.parquet")

# Read with column selection and row filtering (pushdown to file)
subset = pq.read_table(
    "users.parquet",
    columns=["user_id", "revenue"],          # Only read these columns
    filters=[("revenue", ">", 100)],         # Predicate pushdown
)

# Read CSV with type inference
csv_table = pcsv.read_csv("data.csv", convert_options=pcsv.ConvertOptions(
    column_types={"amount": pa.float64(), "count": pa.int32()},
))

# Streaming reads for large files (process in batches)
parquet_file = pq.ParquetFile("large_dataset.parquet")
for batch in parquet_file.iter_batches(batch_size=10_000):
    # Process each batch (RecordBatch) without loading the full file
    filtered = pc.filter(batch, pc.greater(batch["amount"], 0))
    process_batch(filtered)

Zero-Copy Interop

python
# Arrow enables zero-copy conversion between libraries
import pyarrow as pa
import pandas as pd
import polars as pl

# Arrow → Pandas (zero-copy when possible)
arrow_table = pa.table({"x": [1, 2, 3], "y": [4.0, 5.0, 6.0]})
pandas_df = arrow_table.to_pandas()           # Near-instant for compatible types

# Pandas → Arrow
arrow_from_pandas = pa.Table.from_pandas(pandas_df)

# Arrow → Polars (zero-copy)
polars_df = pl.from_arrow(arrow_table)

# Polars → Arrow (zero-copy)
arrow_from_polars = polars_df.to_arrow()

# Arrow enables data exchange between:
# Python ↔ R (via reticulate)
# Python ↔ DuckDB (zero-copy)
# Python ↔ Spark (via PySpark)
# JavaScript ↔ WASM modules

Partitioned Datasets

python
# Work with partitioned datasets on disk or cloud storage
import pyarrow.dataset as ds

# Read a partitioned Parquet dataset (Hive-style partitioning)
# data/
#   year=2025/month=01/part-0.parquet
#   year=2025/month=02/part-0.parquet
#   year=2026/month=01/part-0.parquet

dataset = ds.dataset(
    "s3://my-bucket/events/",
    format="parquet",
    partitioning=ds.partitioning(
        pa.schema([
            ("year", pa.int32()),
            ("month", pa.int32()),
        ]),
        flavor="hive",
    ),
)

# Scan with partition pruning (only reads relevant files)
scanner = dataset.scanner(
    columns=["event_type", "user_id", "timestamp"],
    filter=(ds.field("year") == 2026) & (ds.field("month") >= 1),
)
table = scanner.to_table()

# Write partitioned dataset
ds.write_dataset(
    table,
    "output/events/",
    format="parquet",
    partitioning=ds.partitioning(
        pa.schema([("year", pa.int32()), ("month", pa.int32())]),
        flavor="hive",
    ),
    existing_data_behavior="overwrite_or_ignore",
)

Arrow IPC (Inter-Process Communication)

python
# Share data between processes without serialization overhead
import pyarrow as pa
import pyarrow.ipc as ipc

# Write Arrow IPC format (for streaming between processes)
table = pa.table({"id": [1, 2, 3], "value": [10.0, 20.0, 30.0]})

# File format (random access)
with pa.OSFile("data.arrow", "wb") as f:
    writer = ipc.new_file(f, table.schema)
    writer.write_table(table)
    writer.close()

# Stream format (append-only, lower overhead)
sink = pa.BufferOutputStream()
writer = ipc.new_stream(sink, table.schema)
writer.write_table(table)
writer.close()
buffer = sink.getvalue()    # bytes that can be sent over network/pipe

# Read back
reader = ipc.open_file("data.arrow")
loaded = reader.read_all()

JavaScript (Arrow JS)

typescript
// src/data/arrow-client.ts — Read Arrow data in the browser
import { tableFromIPC, tableToIPC } from "apache-arrow";

// Fetch Arrow IPC data from an API
async function fetchArrowData(url: string) {
  const response = await fetch(url);
  const buffer = await response.arrayBuffer();

  // Parse Arrow IPC format (zero-copy in WASM-backed implementations)
  const table = tableFromIPC(new Uint8Array(buffer));

  console.log(`Loaded ${table.numRows} rows, ${table.numCols} columns`);
  console.log("Schema:", table.schema.fields.map((f) => `${f.name}: ${f.type}`));

  // Access columns
  const ids = table.getChild("id");
  const values = table.getChild("value");

  // Iterate rows
  for (const row of table) {
    console.log(row.toJSON());  // { id: 1, value: 10.0 }
  }

  return table;
}

// Send Arrow data to a server
async function sendArrowData(url: string, table: any) {
  const buffer = tableToIPC(table);
  await fetch(url, {
    method: "POST",
    headers: { "Content-Type": "application/vnd.apache.arrow.stream" },
    body: buffer,
  });
}

Installation

bash
# Python
pip install pyarrow

# JavaScript
npm install apache-arrow

# With DuckDB (Arrow-native)
pip install duckdb    # DuckDB uses Arrow internally

Examples

Example 1: Integrating Apache Arrow into an existing application

User request:

Add Apache Arrow to my Next.js app for the AI chat feature. I want streaming responses.

The agent installs the SDK, creates an API route that initializes the Apache Arrow client, configures streaming, selects an appropriate model, and wires up the frontend to consume the stream. It handles error cases and sets up proper environment variable management for the API key.

Example 2: Optimizing zero-copy interop performance

User request:

My Apache Arrow calls are slow and expensive. Help me optimize the setup.

The agent reviews the current implementation, identifies issues (wrong model selection, missing caching, inefficient prompting, no batching), and applies optimizations specific to Apache Arrow's capabilities — adjusting model parameters, adding response caching, and implementing retry logic with exponential backoff.

Guidelines

1. Parquet for storage, Arrow for compute — Write Parquet to disk/S3; use Arrow in-memory for processing
2. Column pruning — Always specify columns= when reading Parquet; reading all columns wastes I/O and memory
3. Predicate pushdown — Use filters= in Parquet reads; the reader skips row groups that don't match
4. Zero-copy when possible — Use to_pandas(self_destruct=True) for large tables; Arrow can transfer memory ownership
5. Batch processing for large files — Use iter_batches() instead of reading entire files into memory
6. IPC for microservices — Arrow IPC is faster than JSON/CSV for data exchange between services
7. Partitioned datasets for scale — Partition by date/category; queries only scan relevant partitions
8. DuckDB for Arrow queries — DuckDB can query Arrow tables directly with zero copy: duckdb.arrow(table).query("SELECT ...")