sas-parquet

Convert SAS/CSV files to Parquet, detect unique keys, and compare datasets — chunk-by-chunk, memory-safe.

Built for working with large SAS datasets (.sas7bdat) that don't fit in memory. Streams through 100+ GB files on a 16 GB laptop without breaking a sweat.

---

Why

If you've ever tried to:

• Convert a 50 GB .sas7bdat file to Parquet and watched your machine OOM
• Compare a SAS-exported Parquet against a "ground truth" Parquet and not know which column is the primary key
• Validate a SAS-to-Python pipeline migration row-by-row without loading both datasets into memory

…this tool is for you.

What's inside

Three composable command-line tools and a Streamlit web UI:

Command	What it does
sas-parquet convert	Stream a .sas7bdat, .csv, .csv.gz, or .parquet file into one or more Parquet chunks. Smart chunk sizing aims for a target MB per file.
sas-parquet detect-keys	Find the smallest combinations of columns whose values are unique. Useful when you don't know the primary key of a dataset.
sas-parquet compare	Compare two Parquet datasets chunk-by-chunk on a composite key. Reports row counts, schema differences, per-column value mismatches with sample rows.
sas-parquet ui	Launch the four-tab Streamlit UI: Convert · Detect Keys · Compare · Full Pipeline.

Install

pip install sas-parquet              # core CLIs
pip install "sas-parquet[ui]"        # adds the Streamlit web UI
pip install "sas-parquet[ui,dev]"    # everything (ruff, pytest, etc.)

Requires Python 3.10+.

Quick start

Convert SAS → Parquet

# SAS → chunked Parquet (~256 MB per chunk)
sas-parquet convert data/big.sas7bdat --mb-size 256

# SAS → single Parquet file
sas-parquet convert data/big.sas7bdat --single-file

# CSV → chunked Parquet
sas-parquet convert data/big.csv --mb-size 256

# Pick specific columns
sas-parquet convert data/big.sas7bdat --columns loan_id period balance

Detect unique composite keys

Don't know which columns form the primary key of your dataset? Let the tool find them:

sas-parquet detect-keys ./data/loans_parquet/ --first-only

[detect] Total rows: 4,562,891
[detect] Auto-selected 12 candidate(s):
   - loan_id, period, balance, status, origination_date, ...
[key-detect] Searching up to size 4 across 12 candidate column(s)...
--- size 1: 12 combo(s)
  ['loan_id'] -> 1,140,723 dup keys
--- size 2: 66 combo(s)
  ['loan_id', 'period'] -> UNIQUE
[detect] Found 1 minimal unique key(s):
   1. ['loan_id', 'period']

Compare two Parquet datasets row-by-row

sas-parquet compare \
    --source-dir ./data/sas_chunks/ \
    --target-path ./data/prod.parquet \
    --keys loan_id period \
    --output-dir ./reports/

A report file is written with:

• Source and target row counts (from Parquet metadata — milliseconds even on TB-scale files)
• Columns present only on one side
• Data type differences
• Per-column value mismatch counts, sorted by severity
• Sample mismatched rows with key + source value + target value
• Per-chunk timing

Launch the web UI

sas-parquet ui
# Opens http://localhost:8501

Four tabs let you run any single step or chain them end-to-end:

• Convert — point at a .sas7bdat / .csv / .csv.gz / .parquet file, set a target MB per chunk, click run.
• Detect Keys — point at a Parquet file or directory, see the minimal unique composite key(s).
• Compare — pick source and target Parquet, multiselect the composite key, get a per-column mismatch report with downloadable text output.
• Full Pipeline — chain all three: SAS → Parquet → key detection → comparison, with an optional key override if you'd rather skip auto-detection.

Runnable examples

The three examples below are self-contained: each one generates its own data, runs the relevant command(s), and shows the expected output. All you need is pip install sas-parquet.

Example 1 — Convert CSV → Parquet

mkdir -p /tmp/sp_demo && cd /tmp/sp_demo

# Generate a 1,000-row CSV
python -c "
import polars as pl
pl.DataFrame({
    'loan_id': list(range(1000)),
    'period':  ['Q1' if i % 2 == 0 else 'Q2' for i in range(1000)],
    'balance': [round(i * 1.5, 2) for i in range(1000)],
    'status':  ['ACTIVE' if i % 3 else 'CLOSED' for i in range(1000)],
}).write_csv('loans.csv')
"

sas-parquet convert loans.csv --mb-size 5

Expected output:

[convert] Input  : /tmp/sp_demo/loans.csv  (csv)
[convert] Output : /tmp/sp_demo/loans  (chunked)
[convert] Sampling for chunk-size estimate...
[convert] Chunk size: ~500,000 rows per chunk
  chunk 000 -> loans_000.parquet (1,000 rows, 0.0 MB)
[convert] Done. 1,000 rows in 1 chunk(s) -> 0.0 MB in 0.2s

Example 2 — Detect a composite key

Generate a dataset where neither loan_id nor period alone is unique, but the pair (loan_id, period) is. The tool should figure that out without hints.

mkdir -p /tmp/sp_demo2 && cd /tmp/sp_demo2

python -c "
import polars as pl
pl.DataFrame({
    'loan_id': [i // 2 for i in range(1000)],                                  # each loan appears in 2 periods
    'period':  ['Q1' if i % 2 == 0 else 'Q2' for i in range(1000)],
    'balance': [round(((i // 2) % 100 + 1) * 100.0, 2) for i in range(1000)],  # cycles every 100 loans
    'status':  ['ACTIVE' if i % 3 else 'CLOSED' for i in range(1000)],
}).write_csv('loans.csv')
"

sas-parquet convert loans.csv --mb-size 5
sas-parquet detect-keys loans/

Expected detect-keys output (abridged):

[detect] Auto-selected 4 candidate(s):
   - loan_id, balance, period, status

[key-detect] Searching up to size 4 across 4 candidate column(s)...
--- size 1: 4 combo(s)
  ['loan_id'] -> 500 dup keys
  ['balance'] -> 100 dup keys
  ['period']  -> 2 dup keys
  ['status']  -> 2 dup keys
--- size 2: 6 combo(s)
  ['loan_id', 'period'] -> UNIQUE
  ...

[detect] Found 1 minimal unique key(s):
   1. ['loan_id', 'period']

Example 3 — Compare two datasets with known mismatches

Mutate three rows in the target and verify the tool catches exactly those mismatches.

cd /tmp/sp_demo2   # reusing the dataset from example 2

# Build a target with 6 deliberate mismatches (3 loans × 2 periods)
python -c "
import polars as pl
df = pl.read_parquet('loans/loans_000.parquet')
df = df.with_columns(
    pl.when(pl.col('loan_id').is_in([10, 20, 30]))
      .then(pl.col('balance') + 0.99)
      .otherwise(pl.col('balance'))
      .alias('balance')
)
df.write_parquet('target.parquet')
"

sas-parquet compare \
    --source-dir loans/ \
    --target-path target.parquet \
    --keys loan_id --keys period \
    --output-dir reports/

Expected output:

[compare] Source chunks : 1 (1,000 rows total)
[compare] Target rows   : 1,000
[compare] Columns to compare : 2
[compare] Chunk 1/1: loans_000.parquet
   rows=1,000  matched=1,000  source_only=0  (0.2s)
[compare] Matched: 1,000 / 1,000
[compare] 1 column(s) have value mismatches.
[compare] Report: /tmp/sp_demo2/reports/compare_loans_<timestamp>.txt

The report file shows the 6 mismatched rows with their key, source value, and target value:

[balance] — first 6 mismatch(es):
   {'loan_id': 10, 'period': 'Q1', 'balance__source': 1100.0, 'balance__target': 1100.99}
   {'loan_id': 10, 'period': 'Q2', 'balance__source': 1100.0, 'balance__target': 1100.99}
   {'loan_id': 20, 'period': 'Q1', 'balance__source': 2100.0, 'balance__target': 2100.99}
   {'loan_id': 20, 'period': 'Q2', 'balance__source': 2100.0, 'balance__target': 2100.99}
   {'loan_id': 30, 'period': 'Q1', 'balance__source': 3100.0, 'balance__target': 3100.99}
   {'loan_id': 30, 'period': 'Q2', 'balance__source': 3100.0, 'balance__target': 3100.99}

Design notes

• Smart chunk sizing — samples 10K rows, writes a probe Parquet, measures compressed bytes/row, then derives a chunk size targeting your --mb-size (default 50 MB).
• Streaming everywhere — uses pyreadstat.read_file_in_chunks, polars.scan_* lazy frames, and pyarrow.ParquetWriter for incremental writes. No data ever fully materializes in memory.
• Chunk-by-chunk comparison — loads one source chunk, semi-joins target down to matching keys (streaming), compares with null-safe equality, frees memory, repeats. Works on multi-TB datasets.
• SAS empty-string handling — SAS exports missing character values as "". The compare tool normalizes these to null before comparing strings, so they don't show up as false mismatches against true null in Parquet.
• Anti-join source-only detection — keys present in source but missing from target are counted and surfaced separately.
• Metadata-only row counts — pyarrow.parquet.read_metadata gives exact row counts in milliseconds, even on TB-scale datasets, without scanning data.
• Schema-drift-aware reading — chunked datasets with column dtype drift (one chunk has a column as Null, another as Float64) are unified at read time.

What this is not

• A full SAS replacement. It reads .sas7bdat. It does not run PROC SQL, macros, or formats. Use SAS Viya or pysas if you need those.
• Spark or Dask. Single-machine tool optimized for "fits on a beefy laptop or VM" workloads. If you have a cluster and 10+ TB of data, use Spark.
• A diff tool for arbitrary files. Source and target must share a composite key. Schema-drift handling exists but is intentionally minimal.

Limitations and known issues

• SAS numerics are Float64. SAS stores all numbers as 8-byte floats. After conversion, columns that are conceptually integers (IDs, counts) arrive as f64. A future release will auto-downcast int-like floats.
• Float equality is exact. Comparisons use null-safe ==. No fuzzy tolerance for floats yet — 1.0 vs 1.0000001 is a mismatch. --rel-tol is planned for 0.2.0.
• pyreadstat quirks on Windows. Some .sas7bdat files have out-of-range file timestamps that crash pyreadstat 1.3.4's internal datetime.fromtimestamp call. We patch around this transparently in the SAS reader.
• Conversion is single-threaded. Polars parallelizes the column-level work, but the chunk loop itself is serial. For 100+ GB files this is usually I/O bound anyway.

Development

git clone https://github.com/kulbshar/sas-parquet.git
cd sas-parquet
python -m venv .venv
source .venv/bin/activate           # or  .venv\Scripts\activate  on Windows
pip install -e ".[ui,dev]"

pytest                               # 105 tests
ruff check src tests

The test suite covers every module end-to-end with synthetic Parquet fixtures, including the Streamlit UI (driven via streamlit.testing.v1.AppTest).

Changelog

See CHANGELOG.md.

License

MIT — see LICENSE.