Designing Data Storage

Comprehensive guide to selecting and using modern data storage formats for analytics and machine learning. Covers six file formats (Parquet, Arrow, Lance, Zarr, Avro, ORC) and three lakehouse table formats (Delta Lake, Apache Iceberg, Apache Hudi).

Table of Contents

• Quick Format Comparison
  • File Formats
  • Lakehouse Table Formats
• When to Use Which?
  • File Formats
  • Lakehouse Table Formats
• Format Selection Matrix
• Detailed Reference Guides
• Code Examples
  • Parquet (Columnar Analytics)
  • Delta Lake (ACID Table Format)
  • Apache Iceberg (Engine-Agnostic)
  • Lance (ML-Native)
  • Zarr (N-Dimensional Arrays)
• Best Practices
  • File Formats
  • Lakehouse Formats
• Compression Codec Comparison
• Related Skills
• References

Quick Format Comparison

File Formats

Format	Type	Best For	Compression	Schema Evolution	Random Access
Parquet	Columnar	Analytics, data lakes	✅ (Snappy, Zstd, LZ4)	✅ (add/drop)	✅ (row groups)
Arrow/Feather	Columnar	In-memory, IPC, ML	✅ (LZ4, Zstd)	Limited	✅ (record batches)
Lance	Columnar	ML pipelines, vectors	✅ (Zstd, LZ4)	✅	✅ (multi-modal)
Zarr	Chunked arrays	ML, geospatial, N-dim	✅ (Blosc, gzip)	✅ (chunks)	✅ (chunk-level)
Avro	Row-based	Streaming, Kafka	✅ (deflate, snappy)	✅ (full schema)	❌ (sequential)
ORC	Columnar	Hive, Hadoop	✅ (ZLIB, Snappy)	Limited	✅ (stripe-level)

Lakehouse Table Formats

Feature	Delta Lake	Apache Iceberg	Apache Hudi
ACID Transactions	✅	✅	✅
Time Travel	✅	✅	✅
Schema Evolution	✅	Advanced (branching)	✅
Primary Ecosystem	Spark/Databricks	Engine-agnostic	Spark (CDC focus)
Write Optimization	Copy-on-write	CoW, Merge-on-Read	CoW, Merge-on-Read
Python API	deltalake (pure), PySpark	pyiceberg (pure)	PySpark only
Best For	Spark ecosystems	Multi-engine analytics	Change data capture

When to Use Which?

File Formats

Choose Parquet when:

• You need broad compatibility (Spark, DuckDB, Polars, pandas)
• Analytics queries with filtering/aggregation
• Data lake storage with partitioning
• Mature ecosystem with best compression support

Choose Arrow/Feather when:

• Zero-copy sharing between processes (IPC)
• Fast serialization/deserialization for ML training
• In-memory format persistence
• Need Arrow ecosystem (Kernel, CUDA, etc.)

Choose Lance when:

• Machine learning pipelines with embeddings/vectors
• Need multi-modal data (text + images + audio + vectors)
• Versioned datasets with Git-like branching
• Cloud-native (S3/GCS) with no metadata catalog required

Choose Zarr when:

• N-dimensional arrays (tensors, satellite imagery, medical scans)
• Chunked, compressed storage for ML
• Parallel reads/writes across chunks
• Cloud-optimized (s3://, gs:// with fsspec)

Choose Avro when:

• Streaming to Kafka/Kinesis
• Schema evolution is critical (backward/forward)
• Row-based access pattern
• Need to serialize objects/records

Choose ORC when:

• Working primarily with Hive/Hadoop
• Hive ACID transactions
• Legacy big data pipelines

Lakehouse Table Formats

Choose Delta Lake when:

• You're in the Spark/Databricks ecosystem
• Need mature tooling with pure-Python deltalake library
• Want simplest time travel and schema evolution

Choose Apache Iceberg when:

• You need engine-agnostic tables (Spark, Trino, Flink, DuckDB)
• Advanced schema branching and partition evolution
• Prefer explicit catalog abstraction

Choose Apache Hudi when:

• You're building CDC pipelines from Kafka/DB logs
• Need upsert/delete support with streaming
• Low-latency incremental processing

Format Selection Matrix

Use Case	Recommended Format	Reason
Data lake analytics	Parquet or Delta Lake	Mature, partitioning, ecosystem
ML training data	Arrow/Feather or Lance	Zero-copy, vector support
Geospatial arrays	Zarr	Chunked, N-dimensional, cloud-optimized
Streaming/Kafka	Avro	Schema evolution, row-based
Legacy Hive	ORC	Compatibility
Feature stores	Lance or Delta Lake	Versioning, vectors
IPC between processes	Arrow IPC or Feather	Zero-copy, fast
Multi-engine analytics	Apache Iceberg	Engine-agnostic, catalog flexibility
CDC pipelines	Apache Hudi	Built for upserts/streaming
Quick exports	Parquet (Zstd)	Good compression/decompression speed

Detailed Reference Guides

File Format References

• references/parquet.md - Deep dive on Parquet layout, writer settings, partitioning patterns
• references/format-selection-guide.md - Comprehensive decision matrix and selection criteria

Lakehouse Table Format References

• references/delta-lake.md - Pure-Python API, PySpark integration, time travel, cloud storage
• references/iceberg.md - Catalog configuration, schema evolution, partition evolution
• references/hudi.md - CDC patterns, upserts, Copy-on-Write vs Merge-on-Read

Code Examples

Parquet (Columnar Analytics)

import polars as pl
import pyarrow.parquet as pq
import pyarrow as pa

# Write with Polars
df = pl.DataFrame({"id": [1, 2, 3], "value": [100.0, 200.0, 150.0]})
df.write_parquet("data.parquet", compression="zstd")

# Write with PyArrow (more control)
table = pa.Table.from_pandas(df)
pq.write_table(
    table,
    "data.parquet",
    compression="ZSTD",
    compression_level=3,
    row_group_size=100000,
    use_dictionary=True
)

# Read with column pruning
df = pl.read_parquet("data.parquet", columns=["id", "value"])

# Dataset scanning with predicate pushdown
lazy_df = pl.scan_parquet("s3://bucket/dataset/**/*.parquet")
result = lazy_df.filter(pl.col("value") > 100).collect()

Delta Lake (ACID Table Format)

from deltalake import DeltaTable, write_deltalake
import pyarrow as pa

# Create Delta table
data = pa.table({
    "id": [1, 2, 3],
    "value": [100.0, 200.0, 150.0],
    "timestamp": ["2024-01-01", "2024-01-02", "2024-01-03"]
})

write_deltalake("data/delta-table", data, mode="overwrite")

# Read with time travel
dt = DeltaTable("data/delta-table")
dt.load_version(1)  # Load specific version
df = dt.to_pandas()

# Upsert (merge)
new_data = pa.table({"id": [2, 4], "value": [250.0, 400.0]})
dt.merge(
    source=new_data,
    predicate="target.id = source.id"
).when_matched_update_all().when_not_matched_insert_all().execute()

Apache Iceberg (Engine-Agnostic)

from pyiceberg.catalog import load_catalog

# Load catalog
catalog = load_catalog("glue", **{
    "type": "glue",
    "s3.region": "us-east-1",
})

# Create and write table
table = catalog.create_table("default.events", schema=schema)
table.append(parquet_data)

# Read with time travel
df = table.scan(
    row_filter="population > 1000000",
    as_of_timestamp="2024-01-01T00:00:00Z"
).to_pandas()

Lance (ML-Native)

import lancedb
import polars as pl

# Create Lance dataset
db = lancedb.connect("./data.lance")
df = pl.DataFrame({
    "id": [1, 2, 3],
    "text": ["Hello", "World", "ML"],
    "vector": [[0.1] * 128, [0.2] * 128, [0.3] * 128]
})

table = db.create_table("my_table", df)

# Vector search
results = table.search([0.1] * 128).limit(5).to_pandas()

Zarr (N-Dimensional Arrays)

import zarr
import numpy as np

# Create chunked, compressed array
z = zarr.open(
    'data.zarr',
    mode='w',
    shape=(1000000, 1000),
    chunks=(10000, 1000),
    dtype='f4',
    compressor=zarr.Blosc(cname='zstd', clevel=3)
)

# Write and read partial (only loads needed chunks)
z[:10000, :] = np.random.rand(10000, 1000).astype('f4')
slice_data = z[5000:6000, :]

Best Practices

File Formats

1. ✅ Default to Parquet - Broadest compatibility, good compression
2. ✅ Use Arrow/Feather for ML staging - Zero-copy between frameworks
3. ✅ Use Lance when vectors are first-class - No separate vector DB needed
4. ✅ Use Zarr for N-dim arrays - Geospatial, video, 3D data
5. ✅ Compress everything - Snappy (fast) or Zstd (balanced)
6. ✅ Partition wisely - By date/region/tenant to enable pruning
7. ❌ Don't use Avro for analytics - No column pruning, row-based
8. ❌ Don't use ORC unless in Hive/Hadoop world

Lakehouse Formats

1. ✅ Use partitions - Partition by date/region for predicate pushdown
2. ✅ Vacuum regularly - Clean up old files to avoid storage bloat
3. ✅ Optimize layouts - Compact small files for better performance
4. ✅ Catalog choice - AWS Glue for AWS, Hive for on-prem, REST for SaaS
5. ✅ Monitor metadata - Table metadata grows with operations
6. ❌ Don't vacuum too aggressively - Keep time travel window
7. ❌ Don't skip OPTIMIZE - Small files hurt performance

Compression Codec Comparison

Codec	Compression Ratio	Speed	Best For
Snappy	Low (~2:1)	⚡⚡⚡ Fast	Fast analytics, default
Zstd	Medium-High (~4:1)	⚡⚡ Fast	General purpose
LZ4	Low-Medium (~2.5:1)	⚡⚡⚡ Very fast	Real-time streaming
Gzip	High (~5:1)	⚡ Slow	Archival, cold storage
Blosc (zstd)	Medium	⚡⚡	Zarr arrays

Related Skills

• @accessing-cloud-storage - S3, GCS, Azure authentication and I/O patterns
• @building-data-pipelines - ETL patterns using these storage formats
• @engineering-ai-pipelines - Vector storage, Lance, embeddings workflows
• @managing-data-catalogs - Hive, Glue, REST catalogs for Iceberg

References

• Apache Parquet
• Apache Arrow
• Lance Format
• Zarr Format
• Apache Avro
• Delta Lake
• Apache Iceberg
• Apache Hudi