A specification and executable conformance suite for tabular data that must survive version control.
A .mdtbl file is a markdown table where a computed column carries its formula
in the header, rows have opaque ids, and nothing is addressed by position. The
point is that stock git merges it correctly, or refuses — never quietly
wrong, with none of this project's tooling installed.
| id | item | qty | unit | total = qty * unit |
| ------ | ------ | --: | ----: | -----------------: |
| r_0001 | widget | 10 | 12.00 | |
key := id
grand := sum(total)
Two branches insert a row far apart. Stock git merge — no driver, no
.gitattributes, nothing installed — merges cleanly, and grand is correct.
The same table with A1-style references merges just as cleanly and reports a
number that is 27% wrong, with no marker anywhere. That difference is the entire
project.
If your table is a list of facts, you probably do not need this. We replayed
7,446 real commits from four public CSV-in-git registries, including 528 real
three-way merges. Neither .mdtbl nor plain CSV ever merged silently wrong.
For an ordinary sorted registry CSV, stock git already does what this format
promises, and none of those maintainers has wanted a computed column in four
years.
The format earns its keep when a table computes — when a merge can leave a total that is quietly wrong, which is the failure a CSV cannot even represent.
Two branches each adding a column conflict badly, across most of the file. That is inherent to one-row-per-line and it is exactly what happens to a CSV. Nothing here fixes it; a column addition rewrites every row either way.
It is not a data-quality tool. Whether a country code is in ISO 3166, whether a URL resolves, whether a date is plausible — none of that is here, and Frictionless does it well. rowspec asks a smaller question: will this table survive being edited by several people over years?
rowspec check data/No migration. On an ordinary .csv that alone refuses a committed conflict
marker — which Python's own csv module parses as valid rows without
complaint — a duplicate column name, a ragged row, invalid UTF-8, and an
invisible character in a column name. A five-line
data/countries.csv.rowspec.json naming the key and the order column adds the
rest. See docs/csv.md, and copy
docs/ci/rowspec-check.yml into
.github/workflows/.
- uses: actions/checkout@v7
- uses: kindspec/rowspec@v0.3.0
with:
paths: dataIt runs both commands, because that is the whole reason it exists: check
applies the structural refusals, eval computes every table and fails on an
unresolved value, and neither subsumes the other. A misspelled column name
or a pasted 1,299.00 is #REF! and correctly not a structural refusal, so
a pipeline with only check is green on a broken total.
The action's own test suite proves that in both directions: one job asserts it
passes a good tree, and another feeds it a well-formed table whose total is
#REF! and asserts the action fails — then runs the same file with
eval: false and asserts it passes, so the first assertion is about eval
and not about the file.
Or copy docs/ci/rowspec-check.yml if you would rather own the workflow.
Run against 1,564 CSVs in three public data repositories it refused 16: two
grouped-header spreadsheet exports, thirteen truncated CDC snapshots in
owid/covid-19-data whose header declares 14 columns for a 6-field row, and
one deliberately broken fixture in iptv-org/database's own test suite. No
false positives.
The conformance suite is the deliverable. It checks out two branches, runs
stock git merge, evaluates the merged file, and asserts on the computed
number. No prior art does this. The spec, the validator and the reference
implementation exist so the suite has something to check.
619 conformance cases two implementations, both passing
112 mutants, all distinct 110 killed, 2 equivalent
0 survived, 0 stale, 0 broken
The second implementation is the point. reference/rowspec_alt/ was written
from SPEC.md alone by an author forbidden to read reference/rowspec/, and it
runs against the same fixture tree in CI on every push. It is the only evidence
that this document defines the format rather than describing one program, and
it has repeatedly been the half that was right: on the last three questions
where the two disagreed, the independent implementation was correct and the
reference was wrong.
The mutation gate is the other half. The suite is only worth its green tick if it can go red, so the implementation is deliberately broken in 112 specific ways. 110 must be caught by a case; the other two carry a recorded claim that no input could distinguish them from the original, and the gate fails the run if a case turns out to catch one anyway. No two mutants produce the same mutated source, and the gate refuses to run if two do. A mutant that survives is reported as a failure, and so is a stale one whose pattern no longer matches the source — because a check that quietly stopped running is the failure this project keeps finding in itself.
Both are kindkit — the shared machinery every kindspec kind runs on. What lives here is the part that knows what a rowspec case means; walking the tree, reading fixtures as exact bytes, splicing mutants and accounting for the verdicts are the kit's, and rowspec is its first consumer.
Neither number is a claim about correctness in general. See docs/rationale.md for what has been measured, and what has been measured and found wanting.
just setup # uv sync
just check # fmt-check + lint
just test # everything under tests/: suite, gate, CSV, CLI, boundary
just conform # the suite alone, against a stock git binary
just mutants # deliberately break the implementation; the suite must notice
just conform-alt # the SECOND implementation, against the same fixture tree
just test-xlsx # the exporter, WITH its optional extra installed
just test-libreoffice # exported workbooks recalculated by LibreOffice
just run FILE # validate an artifact (rowspec check)
just eval FILE # print computed values and FAIL on any #REF!A real git binary is required. The suite's central claim is about what stock
git does, so it uses stock git.
From the source archive (sdist), run the suite through uv, not pip. The
sdist ships the fixture tree and the runner, but the runner needs
kindkit, which is in the dev
dependency group. Dependency groups are not package metadata, so
pip install rowspec-X.tar.gz cannot install it, and
python conformance/run_cases.py then fails with No module named 'kindkit'.
Unpack the archive and run uv sync (or just setup) in it, then
just conform. kindkit is not on PyPI yet. Once it is, a conformance extra
can make pip enough (#47).
pip install rowspec # no xlsx dependency, and none is coming
pip install 'rowspec[xlsx]' # openpyxl, and the exporter that uses it
python -m rowspec_xlsx table.mdtbl table.xlsxThe exporter lives in export/, outside reference/, and that is the
whole design. reference/ is standard-library-only because a dependency there
is a dependency every independent implementation of the format inherits, and an
independent implementation of a table format must not be required to read xlsx.
A test fails if anything under reference/ ever imports outside the standard
library, so the boundary is a check rather than a convention.
It writes one sheet holding a spreadsheet table named rowspec. Computed
columns are live formulas over structured references, such as
rowspec[[#This Row],[qty]], and the declared aggregates sit below the grid as
SUBTOTAL forms. No formula holds a coordinate. The formulas are written to
carry this format's rules rather than the spreadsheet's: a blank operand, text
in a summed column, or the minimum of nothing is an error, not a 0. A #REF!
arrives as a spreadsheet error value, though without the name in its
parentheses.
The workbook computes with spreadsheet arithmetic, which is not this format's.
Tolerant comparison and subtraction can make an if take the other branch
(#88), and left-to-right
addition can cancel a total that §7's exact sum keeps
(#83). The key column is text,
so a formula using it as a number is an error in the workbook. Hidden or
filtered rows drop out of the SUBTOTAL aggregates but not out of count.
What a spreadsheet cannot say, the exporter refuses rather than approximates,
and writes nothing: cumulative, prior and delta; any where predicate;
and a few narrower cases listed in export/rowspec_xlsx/__init__.py.
What is checked is agreement on the conformance tree, not in general.
just test-libreoffice exports every .mdtbl in the conformance tree that
rowspec accepts and the exporter can express. It recalculates each one in
headless LibreOffice and compares every computed cell and aggregate with
rowspec eval. Numbers are compared at 15 significant digits, which is all
LibreOffice saves. CI runs it. Where they disagree, the test pins the value
LibreOffice gives, so a change in either direction fails it. That covers four
sum and avg cases from the tree near 1e308, plus two tables written for
the test: if(c = 0.3, …) over 0.1 + 0.2 (#88), and 1e16 + 1 - 1e16 (#83).
Both kinds of disagreement occur at ordinary magnitudes, so a table outside
the tree can disagree in ways nothing here checks. For numbers that are
exactly rowspec eval's, see the static-values proposal
(#89).
check applies the refusals in §9 — the structural ones, the things that make a
table survive version control. eval computes the table and fails on any
#REF!.
They catch different things and neither subsumes the other. A misspelled
column name, a pasted 1,299.00, or a non-breaking space in a number are all
#REF! under §8 and are correctly not §9 refusals — the file is well-formed,
its total is wrong. check alone reports 0 refused on a table whose total is
wrong, so a CI recipe built on check is green on a broken total.
Run both:
rowspec check . # will this table survive being edited by several people?
rowspec eval . # does it currently say anything false?Per directory, deliberately — see LICENSE. The fixtures are CC0 so they can be vendored into an implementation in any language under any licence.
Standing on Coopy/daff, ClassSheets, Object Spreadsheets, Lotus Improv and org-mode. See NOTICE.