Checking and summarizing a saved CSV at the command line: Miller, DuckDB, csvtk and qsv

Miller is the safest single tool for this job, because it stops on a malformed row by default. If you think in SQL, DuckDB is the close alternative once you declare the columns.

PRACTICAL PICKSPreferredPickOur Preferred PickA checked CSV handoff at the terminal$0

Miller

One free executable that checks, filters and totals a CSV, and by default stops when a row has the wrong number of fields.

Run check first and write to a temporary file: rows before an error are already printed. Its expression syntax takes learning.

Why Miller is our Pick
Newcomer assessment

No Rising Star selected

The newer tools we examined did not establish all the release, platform and checked CSV capabilities needed for this job.

xsv2 continues an existing product; Hawk and Grab do not establish the complete workflow on both Windows and macOS.

Who is this article for? Terminal users with a saved CSV from a source they trust, who want to inspect it, filter rows and total values, then hand on a new file only when the input is sound and the command succeeded.

About 24 min read

What Preferred Pick, Rising Star and Ready to Rock mean

The quick list

Choose Miller for a checked CSV handoff. All four are free; compare malformed-row handling, query syntax and installation.

All four are free, open-source command-line tools with no account or subscription, checked November 7, 2026. Your costs are setup, upkeep and checking files. Licenses: BSD-2-Clause for Miller; MIT for DuckDB, csvtk and qsv. qsv pro is a separate optional graphical product.

The comparison at a glance

Compare what each tool does with a row that has the wrong number of fields, how to check a whole file before writing output, and what to control on your computer.

Miller leads for a checked CSV handoff, followed by DuckDB, csvtk and qsv among the eight tools scored. Some other workflows could not be established from their documentation. All tools remain selectable.

Command-line tools for checking and summarizing a saved CSV
FeatureMillerDuckDB CLIcsvtkqsv
Best fitPreferred Pick: check, filter and total a saved CSV in one terminal pipelineAlso consider: SQL as your daily language, with the columns declaredSingle-executable table pipelines; Miller remains our choice for a checked handoffAlso consider: set bad rows aside or validate against a schema
Documented default on a row with the wrong number of fieldsStops with a header/data length error; rows before it may already be printedReader stops on structural errors (strict_mode on), but automatic detection can misread a messy file; declare columnsRequires equal field counts; skips lines starting with # by default; set -C for your input and leave -I offvalidate stops and exits nonzero; sqlp fails on long rows unless told to truncate; short rows not documented
Whole-file check before you write outputmlr --icsv check (reads all records, prints nothing)read_csv with columns declared; store_rejects saves bad lines to reject tablesParser enforces field counts; confirm comment settings and accept output only after complete successqsv validate: RFC 4180 and UTF-8, optional JSON Schema; --split-ragged writes .valid and .invalid files
InstallOne executable; releases, Homebrew, MacPorts, Chocolatey, conda-forge, Snap; distro packages may be olderOne executable; download, Homebrew or winget; Windows needs the Visual C++ RedistributableOne executable; Linux, macOS and Windows; downloads, Homebrew, conda and PixiPrebuilt downloads, Homebrew, Scoop, own apt repository; x86-64 downloads need AVX2 since 24.0.0; features differ by package
Local behavior to controlReads .mlrrc from the current folder (MLRRC=__none__ turns it off); expressions can run shell commands (--no-shell blocks them)Treats SQL and files as trusted; installs and loads signed extensions on demand by default (can be turned off)Comment lines skipped by default; repair and ignore-illegal-row options can change or discard recordsFull builds include web, geocoding, AI-model and self-update commands
Maintenance and ownershipSteady 2026 releases, mostly by its creator, John KerlDuckDB Foundation holds the project; Amazon agreed in August 2026 to acquire DuckLabs, its developerShen Wei; current releases; 0.39.0 fixes filtering, numeric precision and grouped-summary selectiondatHere, Inc.; “very rapid release tempo”; three breaking major versions between October 5 and 19, 2026
Reader reviewsNo reviews yet. Write a reviewNo reviews yet. Write a reviewNo Practical Picks reader reviews recorded for this comparisonNo reviews yet. Write a review

For saved CSV files from sources you trust. Check the whole file first, write to a temporary file and rename it only after the command succeeds. All four are free under their published licenses, checked November 7, 2026. Miller remains our recommendation for a checked handoff. Other evidence checked October 19, 2026; no tool was run for this review, and the dated October 1 check is a historical record. Next substantive review: December 18, 2026, or sooner after a material parsing or security change. Earlier check record.

All features and sources 13 criteria

Open the groups that matter to you. They compare the same selected tools. “Planned” means future work, not an available feature. Sources include the published conditions.

Checking the input and writing the output

Show 7 criteria
Checking the input and writing the output
FeatureMillerDuckDB CLIcsvtkqsv
Best fitPreferred Pick: check, filter and total a saved CSV in one terminal pipelineAlso consider: SQL as your daily language, with the columns declaredSingle-executable table pipelines; Miller remains our choice for a checked handoffAlso consider: set bad rows aside or validate against a schema
Documented default on a row with the wrong number of fieldsStops with a header/data length error; rows before it may already be printedReader stops on structural errors (strict_mode on), but automatic detection can misread a messy file; declare columnsRequires equal field counts; skips lines starting with # by default; set -C for your input and leave -I offvalidate stops and exits nonzero; sqlp fails on long rows unless told to truncate; short rows not documented
Whole-file check before you write outputmlr --icsv check (reads all records, prints nothing)read_csv with columns declared; store_rejects saves bad lines to reject tablesParser enforces field counts; confirm comment settings and accept output only after complete successqsv validate: RFC 4180 and UTF-8, optional JSON Schema; --split-ragged writes .valid and .invalid files
Earlier hands-on check, October 1, 2026 (before our research-only method)6.22.0: check exited 1 on the short row; a direct transform printed earlier rows, then failed1.5.6: automatic read returned one whole-line column, no error; declared columns failedNot included in the October 1 check23.0.1: validate exited 1; sqlp alone padded the short row and reported success
Inspecting the filehead, summary and describe (added in 6.20.2); pretty-printed outputDESCRIBE and SUMMARIZE in SQLhead, headers, dim, nrow and ncolheaders, stats, frequency and many other commands
How you filter and total by groupfilter then stats1 in Miller’s own expression syntaxSQL: SELECT … GROUP BYfilter or filter2, then summary grouped by selected fieldsSQL through sqlp, or stats and frequency commands
Writing the handoff fileWrite to a temporary file; rename only after success, because earlier rows print before an errorCOPY … TO a file; declare columns first so a bad file failsSet comment handling deliberately; write to a temporary file and rename only after the whole pipeline succeedsValidate first; sqlp --output writes the result
Sources for checking the input and writing the output

csvtk input rules and commands and DuckDB CSV options, checked November 7, 2026. Other documentation checked October 19, 2026: Miller record heterogeneity and verbs; DuckDB CSV options, auto detection and CSV reader write-up; qsv validate and SQL; csvkit csvclean. The dated row records one small Linux check on October 1, 2026, made before our research-only method; it is a historical observation, not current testing. Earlier check record, October 1, 2026.

Installation, maintenance and risk

Show 6 criteria
Installation, maintenance and risk
FeatureMillerDuckDB CLIcsvtkqsv
License and costFree; BSD-2-ClauseFree; MITFree; MITFree; MIT; separate optional qsv pro app
Release evidence (dates shown per tool)6.22.0, September 25, 20262.0.0, October 21, 2026; checked November 7, 20260.39.0, November 5, 2026; checked November 7, 202626.0.0, October 19, 2026, after 24.0.0 and 25.0.0 earlier the same month
InstallOne executable; releases, Homebrew, MacPorts, Chocolatey, conda-forge, Snap; distro packages may be olderOne executable; download, Homebrew or winget; Windows needs the Visual C++ RedistributableOne executable; Linux, macOS and Windows; downloads, Homebrew, conda and PixiPrebuilt downloads, Homebrew, Scoop, own apt repository; x86-64 downloads need AVX2 since 24.0.0; features differ by package
Local behavior to controlReads .mlrrc from the current folder (MLRRC=__none__ turns it off); expressions can run shell commands (--no-shell blocks them)Treats SQL and files as trusted; installs and loads signed extensions on demand by default (can be turned off)Comment lines skipped by default; repair and ignore-illegal-row options can change or discard recordsFull builds include web, geocoding, AI-model and self-update commands
Published security recordOne advisory, 2020 (.mlrrc code execution, fixed in 5.9.1); no security policyEngine advisories fixed in 1.1.0 and 1.4.2; npm packages compromised in 2025 (not the CLI); AWS extension advisory, August 2026No published advisories; no security policy; checked November 7, 2026No published advisories; security policy; 2026 fixes for wrong counts, malformed-CSV undefined behavior and cache-file permissions
Maintenance and ownershipSteady 2026 releases, mostly by its creator, John KerlDuckDB Foundation holds the project; Amazon agreed in August 2026 to acquire DuckLabs, its developerShen Wei; current releases; 0.39.0 fixes filtering, numeric precision and grouped-summary selectiondatHere, Inc.; “very rapid release tempo”; three breaking major versions between October 5 and 19, 2026
Sources for installation, maintenance and risk

csvtk releases and security page; DuckDB 2.0 release and settings, checked November 7, 2026. All four licenses checked November 7, 2026. Other evidence checked October 19, 2026: Miller releases, .mlrrc rules and security page; DuckDB release calendar, security policy, settings and Amazon announcement; qsv changelog, README and security policy; csvkit changelog and package record. A published record lists reported issues; it does not prove a tool is free of defects.

What matters when choosing

Miller is our Preferred Pick if you work at a terminal and need to check, filter and total a saved CSV before you hand a new file to someone else. It is a free, BSD-licensed single program for Linux, macOS and Windows. Its documented default is the behavior this job needs: when a row has more or fewer fields than the header, Miller stops with a “header/data length mismatch” error rather than padding or skipping the row, and its check verb reads the whole file without printing anything. Filtering, group totals and CSV output then happen in one chained command. The October 19 evidence covers version 6.22.0, released September 25, 2026. Record heterogeneity, verbs reference, 6.22.0 release

On five criteria we fixed before scoring, Miller scored 4.03 out of 5 and the DuckDB command line 3.75. DuckDB scored higher for everyday use; Miller’s lead comes from its stricter default for malformed rows and fewer local risks. Two habits matter with Miller. It prints the rows it has already read before it reaches a bad one, so run mlr --icsv check first and write the result to a temporary file that you rename only after the command succeeds. It also reads a .mlrrc settings file from the current folder, so set MLRRC=__none__ in scripts to ignore settings you did not write. Customization and .mlrrc

If SQL is your daily language, consider the DuckDB command line. It queries a CSV directly, and its reader stops at structural errors by default. Automatic detection, however, guesses the file’s layout from a sample, and DuckDB’s own 2025 write-up shows it misreading a messy file without an error. Declare the columns so that a malformed file fails. DuckDB’s security policy says it treats both your SQL and your data files as trusted, and by default it downloads and loads signed extensions when a query needs one; its documentation shows how to turn that off. Amazon agreed in August 2026 to buy DuckLabs, the company that develops DuckDB; the project stays with the nonprofit DuckDB Foundation under the MIT license. Version 2.0 was released October 21, 2026; its documented strict parsing and automatic extension settings remain the same. Releases, checked November 7, 2026 CSV options, CSV reader write-up, security policy, Amazon announcement

If bad rows must be set aside rather than the whole file rejected, consider qsv. Its validate command checks structure and UTF-8 encoding, and since August 2026 its --split-ragged option writes good and bad rows to separate files while still failing the check. Pin the version: qsv released three major versions, each with breaking changes, between October 5 and October 19, 2026, and its x86-64 downloads now need a processor with AVX2. csvtk is another single executable with filtering and grouped summaries, but it skips lines starting with # unless you change that setting. csvtk documentation, checked November 7, 2026 qsv validate, qsv changelog, csvclean

No Rising Star is selected: none of the newer tools we examined established the release date, platform support and complete checked CSV workflow needed for this recommendation. These are researched judgments from documentation, release records, issue trackers and security advisories checked October 19, 2026, with prices, additional candidates and DuckDB’s release checked November 7, 2026. An earlier hands-on check on October 1, 2026, before we adopted our research-only method, matched the documented behavior described here; it used one small file and is not the basis of the ranking. This comparison is about files from sources you trust; DuckDB’s security policy says not to open data files from untrusted parties.

Disclosure: Practical Picks articles are researched and written with AI, currently Anthropic’s Claude. Miller, our Pick, ships an agent skill and an MCP server for Claude Code, and qsv ships a Claude Cowork plugin and an MCP server. We have no relationship with either project or with Anthropic over this article, and these integrations carried no weight in our judgment. Miller agent skill, qsv

Miller

Preferred Pick · a checked CSV handoff at the terminal

$0BSD-2-Clause · single executable

Choose Miller to check, filter and total a saved CSV in one command, with a strict default when a row has the wrong number of fields. Run check first and rename the output only after success.

Reasons to choose it

  • Documented default stops on a row with the wrong number of fields
  • check verb reads the whole file without printing output
  • One executable with broad packaging and steady 2026 releases

What to watch

  • Rows before an error are already printed, so stage the output
  • Reads .mlrrc from the current folder unless MLRRC=__none__
  • Its own expression syntax takes learning

Miller reads CSV with --icsv, writes it with --ocsv and chains verbs with then, so one command can check, select fields, filter rows, total them by group and sort the result. Its verbs include check, head, cut, filter, stats1, sort and summary, and version 6.20.2 (July 2026) added describe, which shows a compact schema of the input. It also reads and writes TSV, JSON and several other formats. Miller is one executable under the BSD-2-Clause license. You can install it from the project’s releases, Homebrew, MacPorts, Chocolatey, conda-forge or Snap, where the command is named miller; the project warns that apt and yum packages may be older, so check mlr --version. Verbs reference, installation, license

Its handling of a malformed row is the main reason for our pick. Miller’s documentation shows that, by default, a row with a different number of fields from the header stops the run with an error naming the file and row, after the records before it have already been printed. The --allow-ragged-csv-input flag fills short rows instead; leave it off for this job. Version 6.21.0 (August 10, 2026) fixed a race that could make exit codes unreliable when a transformation failed, so use a current release. In an earlier hands-on check on October 1, 2026, before we adopted our research-only method, Miller 6.22.0’s check exited with status 1 on a file with one short row, and a direct transformation of that file printed earlier rows before failing. Record heterogeneity, 6.21.0 release, earlier check record, October 1, 2026

Read the full Miller analysis and sources

Control two local behaviors in scripts. Miller reads .mlrrc settings from your home folder, your configuration folder and then the current folder, so a folder you receive can change how Miller parses files; MLRRC=__none__ turns all of them off. Its only published security advisory, from 2020, was code execution through a .mlrrc in the working directory, fixed in 5.9.1 by refusing commands such as --prepipe in those files. Miller’s expression language can also run shell commands through its system and exec functions; since 6.20.2, --no-shell or MLR_NO_SHELL=true blocks them. Customization, 2020 advisory, shell gate

Miller has released steadily in 2026, in February, June, July, August and September, mostly through its creator, John Kerl, and it publishes no security policy. A 2025 report of fields dropped from ragged, headerless TSV and CSV-lite input was fixed in 6.19.0 in June 2026; it did not affect --icsv. Its tradeoff is syntax: expressions such as filter '$amount >= 8' need shell quoting and some learning. Choose DuckDB if you would rather write SQL. Issue 1749, 6.19.0 release, 6.20.2 release

Used Miller? Share your experience to help other readers choose.

DuckDB CLI

Also consider · SQL as your daily language

$0MIT · single executable

Choose the DuckDB command line if you think in SQL. Declare the columns so a malformed file fails, and turn off automatic extension installs for local files.

Reasons to choose it

  • Queries a CSV in place with SQL
  • Strict CSV reader once the layout is declared
  • Reject tables can keep bad lines for review

What to watch

  • Automatic detection can misread a messy file without an error
  • Treats SQL and data files as trusted; installs extensions on demand by default
  • Version 2.0 released October 21, 2026; DuckLabs acquisition closing unconfirmed

DuckDB is an analytical database engine, and its command line is a single program for Linux, macOS and Windows, available as a direct download, through Homebrew or through winget; on Windows it needs the Microsoft Visual C++ Redistributable. You can query a CSV in place with read_csv, look at it with DESCRIBE or SUMMARIZE, and write results with COPY ... TO. It is MIT-licensed. Version 2.0.0 was released October 21, 2026, according to the release page checked November 7, 2026. The October 19 release calendar listed support for the 1.5 series until March 1, 2027. Releases, install, release calendar, 1.5.6 release

Its CSV reader is strict once it has the right layout: strict_mode is on by default, the scan stops at the first structural error, and padding missing columns or ignoring errors are opt-in. store_rejects saves bad lines to reject tables instead. The risk is the guess that comes first. Automatic detection samples 20,480 rows and picks the delimiter and quoting that give the most consistent column count. DuckDB’s April 2025 write-up shows it deciding that a messy file had no quotes and skipping its header rows, which produced a wrong table; in full automatic mode it read about 91% of the data in a benchmark of messy files correctly. Passing columns, which turns detection off, together with header and delim, makes a malformed file fail. In an earlier hands-on check on October 1, 2026, before our research-only method, DuckDB 1.5.6’s automatic read of a file with one short row returned a single whole-line column without an error; with the four columns declared, it failed. CSV options, auto detection, faulty files, CSV reader write-up, earlier check record, October 1, 2026

Read the full DuckDB CLI analysis and sources

Treat it as a powerful local engine, not a sandbox. DuckDB’s security policy says SQL is executable code, that its file readers assume “a well-formed file from a trusted writer”, and “Do not open data files from untrusted parties.” By default, DuckDB installs and loads a core extension when a query needs one, for example to read a CSV from a web address; core extensions come from the project’s repository and are checked against built-in signing keys before loading. For local files, SET autoinstall_known_extensions = false; and SET autoload_known_extensions = false; turn that off, and allowed_directories with enable_external_access = false limits which folders it can read. Its safe mode blocks reading CSV files altogether, so it does not suit this job. Security policy, securing DuckDB, extensions, settings, signing, safe mode

Its security record is published. A 2024 flaw let sniff_csv read files when external access was disabled (fixed in 1.1.0), and 2025 weaknesses in database encryption were fixed in 1.4.2. In September 2025 attackers briefly published malware in DuckDB’s Node.js packages on npm, not in the command line. An August 2026 advisory concerned the separate AWS extension, which this job does not use. Security advisories, npm incident, AWS extension advisory

Amazon announced an agreement to acquire DuckLabs in late August 2026; Amazon says it is not acquiring the open-source project, which stays with the DuckDB Foundation under the MIT license, and that the founders continue to lead its technical direction. Techzine reported that it was set to take effect in early September 2026; we did not find a primary confirmation that it has closed. Choose Miller if you want a strict check without declaring a schema first. Amazon announcement, Techzine, August 26, 2026

Used DuckDB CLI? Share your experience to help other readers choose.

csvtk

Single-executable table pipelines

$0MIT · single executable

Choose csvtk if its separate table commands suit your scripts. Set comment handling deliberately and leave row-skipping options off. Miller remains our preference for a dedicated whole-file check.

Reasons to choose it

  • Parser requires equal field counts
  • Filtering and grouped summaries
  • One executable for Linux, macOS and Windows

What to watch

  • Skips lines starting with # by default
  • Repair and row-skipping flags can discard or alter data
  • Current release fixes filtering and grouped-summary correctness issues

csvtk inspects tables, selects columns, filters rows and computes grouped summaries. Its documentation lists Linux, macOS and Windows support, downloadable binaries and package-manager installation. It is MIT-licensed and free. These are documented capabilities, not measurements of speed. Repository and documentation, checked November 7, 2026

Its parser requires the same number of fields in each row; -I/--ignore-illegal-row skips invalid rows, so leave that option off for a checked handoff. Another default needs attention: it ignores lines starting with #. The documentation advises changing -C to another uncommon symbol when the header starts with #; choose that setting deliberately for your input so data is not silently lost. Write output to a temporary file and accept it only after the complete command succeeds. Miller remains our preference for a dedicated whole-file check and fewer parsing settings to get right. Input rules, checked November 7, 2026

Read the full csvtk analysis and sources

Version 0.39.0, released November 5, 2026, fixes numeric filtering, expression precision and grouped-summary selection, among other issues. Its security page lists no published advisories and no security policy; that does not establish freedom from defects. Choose csvtk if its separate table commands suit your existing scripts. Our score is 3.68 out of 5, behind Miller’s 4.03 and DuckDB’s 3.75. Releases, security page, checked November 7, 2026.

Used csvtk? Share your experience to help other readers choose.

qsv

Also consider · set bad rows aside

$0MIT · prebuilt downloads; optional separate qsv pro

Choose qsv when bad rows should go to a separate file for review, or when you validate against a schema. Validate before any query and pin the version.

Reasons to choose it

  • Validator checks structure, UTF-8 and optional JSON Schema
  • --split-ragged writes good and bad rows to separate files
  • SQL, statistics and many other commands in one toolkit

What to watch

  • Three breaking major versions between October 5 and 19, 2026
  • x86-64 downloads need a processor with AVX2
  • Full builds include web, geocoding and AI-model commands

qsv is a Rust toolkit from datHere, Inc., forked from xsv in September 2021, with dozens of commands, including validate, stats, frequency, select, search and sqlp, which runs SQL through the Polars engine. It is MIT-licensed and ships as prebuilt downloads and through Homebrew, Scoop and the project’s own apt repository; Homebrew’s build enables extra commands such as web fetching and geocoding. Since 24.0.0 (October 5, 2026), the x86-64 downloads require a processor with AVX2, and older processors or conservatively configured virtual machines fail with “Illegal instruction”. A qsvlite build, about 16% of the full size, drops most extras; the full Linux executable downloaded for our October 1 check was about 186 MB. README, 24.0.0 release

Its validator is the most complete here. Without a schema, validate checks that the file follows the RFC 4180 CSV standard and is UTF-8, and exits nonzero if not; with a JSON Schema it also checks values. Since 22.0.1 (August 9, 2026), --split-ragged sends rows with the wrong number of fields to an .invalid file and the rest to a .valid file, still exiting nonzero; malformed quoting and non-UTF-8 data still stop it. Run it before a query. sqlp fails on rows with too many fields unless told to truncate them, but its help does not say what happens to rows with too few. In an earlier hands-on check on October 1, 2026, before our research-only method, qsv 23.0.1’s sqlp padded a short row with empty cells and reported success. Validate, SQL, earlier check record, October 1, 2026

Read the full qsv analysis and sources

The reason to pin a version is change. Versions 24.0.0, 25.0.0 and 26.0.0 arrived on October 5, 14 and 19, 2026, each with breaking changes; in 26.0.0, generated schemas became stricter and a delimiter setting started to affect more output. The project’s own correctness work in September and October 2026 fixed a failed index that made count report a wrong number without an error, undefined behavior in stats and frequency on malformed CSV, and stats cache files that could be readable by other users of the computer. Its security policy describes a “very rapid release tempo” and lists no published advisories. Full builds also include commands that call web services, geocoders and AI models, and a self-updater; that is useful, but it is more to control on a computer that handles private files. qsv pro is a separate graphical product. Changelog, 26.0.0 release, count issue, security policy

Choose qsv when the job is to keep good rows moving and hand bad ones back for review, or to validate against a schema. Choose Miller when the whole file should be rejected and you want fewer moving parts.

Used qsv? Share your experience to help other readers choose.

Which product are we missing?

Know a product that belongs in this article? Tell us what it does differently and why it deserves a closer look. Newcomers welcome.

Readers and product teams can suggest candidates. Suggestions inform our research; they do not guarantee inclusion or a recommendation.

Suggest a product

Opens an email draft with this article attached as a link. Nothing is sent until you send it.

No email app? Send the product name, website, and this article’s title to admin@practicalpicks.guide.

What people are saying

We have no Practical Picks reader reviews for these tools yet. Public reports point the same way as the documentation: checks need to be explicit. In a July 2026 qsv discussion, a user building a validation script asked how to set ragged rows aside without changing them; another user noted that validate stopped at the first bad row, and the maintainer said he would add a way to split them. qsv 22.0.1 added --split-ragged in August. A February 2025 Miller report described fields dropped from ragged, headerless TSV input in a mode this job does not use; it was fixed in June 2026. A September 2026 report in qsv’s tracker showed count returning a wrong number with a success status after a failed index; it was fixed within two days. These are examples, not measures of how often problems occur. qsv discussion, Miller issue 1749, qsv issue 4615

How we researched

This is public-source research. For this review we did not install or run any tool, open any account or contact any project. No project supplied access, payment or input, and our links carry no affiliate tracking. We first compared Miller, DuckDB, qsv and csvkit on October 1, 2026, and rechecked their material claims on October 19, 2026: releases, documented defaults for malformed rows, install routes, licenses, security policies and advisories, issue trackers, ownership and maintenance. We read project documentation and changelogs, GitHub release and issue records, the OSV vulnerability database, PyPI package records and independent press.

Before we adopted our research-only method on October 3, 2026, we ran an earlier hands-on check on October 1, 2026: Miller 6.22.0, qsv 23.0.1, csvkit 2.2.0 and DuckDB CLI 1.5.6 on Linux, against one small CSV with quoted commas and accented names and a copy with one short row. Those four tools produced the expected totals from the clean file; their short-row results are documented in the historical record. It was not a speed, scale, security or cross-platform check. We keep it as a dated historical record, not current testing. Earlier check record, October 1, 2026; clean CSV; CSV with a short row.

We wrote this for someone at a terminal with a saved CSV from a source they trust, who wants to inspect it, filter rows, total values by group and pass on a new file only when the input is sound and the command succeeded. Before scoring, we fixed five criteria and their weights: fit for that job (25%), safety with malformed input (25%), routine use and documentation (15%), setup and portability (15%), and maintenance, change and local risk (20%). Miller scored 4.03 and DuckDB 3.75, followed by csvtk at 3.68 and qsv at 3.48. Miller stayed ahead when we gave more weight to malformed input, to everyday use or to every criterion equally. For a reader who writes SQL every day and weights everyday use heavily, DuckDB came out ahead, which is why we name it as the alternative.

We revisited the alternatives on November 7, 2026, using the CSV directory, recent Hacker News launches and project records. Miller remains our recommendation for rejecting a malformed file before handoff. The four tools above lead among the options we could fully assess. csvkit scored 3.28: it remains useful for teams that already manage Python, but its checks are opt-in and its dependencies add upkeep. Run csvclean --length-mismatch before querying. xan scored 3.35: its documentation expects well-formed input and recommends repair for unequal rows; its February 2024 first release also puts it outside our two-year newcomer window. CSV directory, recent launches, csvkit, xan, checked November 7, 2026.

GoAWK scored 3.45 and GNU awk 3.30. Both can filter, accumulate totals and write quoted CSV, but you must program the field-count check and failure handling yourself; GNU awk also needs the CSV-output helper from its manual. Neither offers Miller’s dedicated check command. sq’s type inference and CSV caching leave questions about preserving values and rejecting malformed rows. zsv documents a check command, but we could not establish its complete rejection behavior. Nushell and Frictionless document CSV parsing or transformation pipelines, but the material we read does not establish the full checked export. We could not establish clickhouse-local’s complete workflow from its available documentation. GoAWK CSV manual, GNU awk manual, sq CSV driver, zsv, Nushell CSV parser, Frictionless transformations, checked November 7, 2026.

Our November 7, 2026 newcomer search found no qualifying Rising Star. xsv2 describes a modernization of xsv, which does not restart the product’s release date. Hawk demonstrates grouped CSV analysis, but its earliest release, Windows support, CSV export and structural rejection are not established. Grab documents extraction into JSON for other tools, rather than the complete grouped CSV handoff. Tennis could not be identified from the search results. QueryPad is a browser SQL playground and Datacmd builds dashboards. Unknown release dates prevent a newcomer award; they do not remove a tool from consideration. xsv2, Hawk, Grab, checked November 7, 2026.

We continue to set aside archived xsv and inactive csvq, q and dsq on the October 19 maintenance evidence. VisiData is an interactive companion, and sqlite3’s import pads short rows with NULL rather than blocking the handoff. tsv-utils replaces embedded tabs and newlines when converting CSV to TSV, which does not preserve arbitrary CSV values for this job. PowerShell’s Import-Csv documentation says extra row values are ignored. csvtool’s row operations and the TextQL and RBQL query descriptions do not establish a complete structural gate and lossless grouped export. CSVfix’s project details and price could not be confirmed. tsv-utils conversion rules, PowerShell Import-Csv, csvtool source, TextQL, RBQL, checked November 7, 2026.

ToolPublic sourcesWhat remains uncertain
MillerDocumentation on malformed rows, verbs, settings files, installation and the MCP server; releases 6.19.0 to 6.22.0; license; security advisory; issue tracker.Exit status of check is not printed in the documentation; behavior on very large or unusual files; continuity with one main maintainer.
DuckDB CLICSV, detection, faulty-file, security, extension, settings and safe-mode documentation; security policy and advisories; CSV reader write-up; release calendar; Amazon’s announcement and press coverage.How often detection misreads ordinary files; governance after the DuckLabs acquisition. Version 2.0 and CSV/settings documentation checked November 7, 2026.
qsvREADME, validate and SQL help, changelog, releases 24.0.0 to 26.0.0, security policy, issue and discussion records.How sqlp treats short rows in current versions; stability of options under its release pace; binary size and CPU needs on your machine.
csvtkRepository, input rules, 0.39.0 release and security page, checked November 7, 2026.Complete failure behavior across commands; effects of comment settings on your input.
Newer alternativesNovember 7, 2026 launch and directory searches, with the October 19 records.xsv2 continues an existing product; Hawk lacks confirmed release, platform and full-workflow evidence; Grab does not show the full job; Tennis is unidentified. None qualifies for the newcomer award.

None of this establishes speed, scale, crash safety, behavior on files from untrusted sources or results on another operating system. All four are free under their published licenses, checked November 7, 2026; your costs are setup, upkeep and checking. Next substantive review is due December 18, 2026, or sooner after a material parsing or security change. Found a changed default or release? Contact the editors.

Sources and evidence

Sources were checked October 19, 2026 unless another date is given. November 7, 2026 checks cover licenses, candidate documentation and DuckDB’s release, CSV options, settings and security page. Project material establishes documented behavior and published releases, not behavior on your files. The October 1, 2026 check is a dated historical observation made before our research-only method.

No screenshots are shown. This comparison rests on public documentation and records; an earlier check on October 1, 2026, before our research-only method, used original synthetic CSV files. Earlier check record, October 1, 2026.