dsv — Data Swiss Knife

A fast CSV, TSV, Parquet, and JSONL command line toolkit written in Rust. Modern fork of BurntSushi/xsv (10.8K ★), merging all 33 unmerged PRs with Parquet/JSONL support, new commands, and modern dependencies.

Install

cargo install dsv

Or download a pre-built binary.

Commands

Quick examples

# Stats on a CSV
dsv stats data.csv --everything | dsv table

# Convert CSV to Parquet
dsv convert data.csv data.parquet

# Convert Parquet back to JSONL
dsv convert data.parquet data.jsonl

# Run stats directly on Parquet (no intermediate CSV)
dsv stats data.parquet

# Search CSV, pipe through table
dsv search --regex '\d{5}' data.csv | dsv table

# Forward-fill empty cells in a column
dsv fill --select Population data.csv

# Rename columns
dsv rename 'old_name:new_name' data.csv

# Transpose rows and columns
dsv transpose data.csv

Why dsv over xsv, qsv, or xan?

dsv fills the gap between the archived xsv and the bloated forks:

xsv was archived in 2023. qsv adds heavy SQL, networking, and pandas-like features — great for ETL but not a drop-in xsv replacement. xan focuses on data journalism with Python integrations. dsv is the lean, modern evolution of xsv: same philosophy, more formats, community features merged.

Feature parity with xsv

Every xsv command works exactly as documented upstream — same flags, same output format — so dsv is a drop-in replacement:

In addition, all 33 community PRs that sat unmerged in xsv's upstream repository are merged here, and every command also accepts Parquet and JSONL input where it makes sense (e.g. stats, select, search).

Whirlwind tour

# Download sample data
curl -LO https://burntsushi.net/stuff/worldcitiespop.csv

# Examine headers
dsv headers worldcitiespop.csv

# Get column statistics
dsv stats worldcitiespop.csv --everything | dsv table

# Create an index for faster access
dsv index worldcitiespop.csv

# Select columns and sample 10 random rows
dsv select Country,AccentCity,Population worldcitiespop.csv \
  | dsv sample 10 \
  | dsv table

# Frequency table
dsv frequency worldcitiespop.csv --limit 5

# Search for rows with population data
dsv search -s Population '[0-9]' worldcitiespop.csv \
  | dsv select Country,AccentCity,Population \
  | dsv table

# Convert to Parquet for analytics
dsv convert worldcitiespop.csv worldcitiespop.parquet

# Same stats, now on Parquet
dsv stats worldcitiespop.parquet --everything | dsv table

JSONL notes

• Only .jsonl and .ndjson extensions are detected as JSONL. Plain .json files are treated as CSV — JSON (non line-delimited) files are not supported as input.
• JSON objects are unordered maps. When reading JSONL, columns are derived from the first record's keys; records with extra keys extend the column set. When writing JSONL, keys are serialized in the order the columns were read (CSV header order, or insertion order for JSONL→JSONL).
• To round-trip reliably between CSV and JSONL, keep column names unique and stable. Empty and non-UTF-8 fields are handled as JSON null.

Performance

dsv inherits xsv's zero-compromise performance model: constant-time indexing, streaming row processing, and minimal memory overhead. On the exact dataset used in xsv's original benchmarks, dsv beats upstream xsv on the same hardware on sort, numeric sort, frequency, and stats (up to 48% faster), and ties on count/select. Parquet reads leverage Apache Arrow's columnar format for fast predicate pushdown and vectorized decoding.

Benchmarks are maintained at BENCHMARKS.md.

Motivation

When handed a 40GB CSV file, existing tools were too slow or inflexible. xsv solved this with indexing, slicing, and composable commands. dsv extends that foundation to modern data formats (Parquet, JSONL) while merging community contributions that sat unmerged for years. If you work with tabular data on the command line, dsv is the tool xsv should have become.

License

Dual-licensed under MIT or UNLICENSE.

dsv is a fork of BurntSushi/xsv by Andrew Gallant (BurntSushi).