Most "I just need a tiny database" problems get solved badly: a CSV that grows types it can't enforce, a SQLite file you can't diff in git, or a YAML blob you end up grepping with regex. GNU recutils sits in the gap. It's a 2009-era GNU project that defines a plain-text record format with types, keys, foreign keys, and constraints — plus a family of CLI tools to query and mutate it. Files are human-readable, vim-editable, git-diffable, and queryable with real expressions.

A .rec file looks like this:

%rec: Book
%type: Year int
%type: Rating range 1 5
%mandatory: Title
%key: Id

Id: 1
Title: The C Programming Language
Author: Kernighan, Ritchie
Year: 1978
Rating: 5

Id: 2
Title: The UNIX Programming Environment
Author: Kernighan, Pike
Year: 1984
Rating: 5

Now query it with proper expressions, not awk gymnastics:

# SELECT * WHERE Year > 1980
recsel -t Book -e "Year > 1980" books.rec

# SELECT Title, Author WHERE Author matches regex
recsel -t Book -p "Title,Author" -e "Author ~ 'Kernighan'" books.rec

# GROUP BY Author with count
recsel -t Book -G Author -p "Author,Count(Id)" books.rec

# Sort, limit, sum
recsel -t Book -S Year -n 3 books.rec
recsel -t Book -p "Sum(Rating)" books.rec

Mutations happen in place, and the tools refuse invalid writes:

recins  -t Book -f Id -v 3 -f Title -v "K&R 2nd Ed" -f Year -v 1988 books.rec
recset  -t Book -e "Id = 1" -f Rating -s 4 books.rec
recdel  -t Book -e "Id = 2" books.rec
recfix  --check books.rec   # validate types, keys, mandatory fields

Multi-table with foreign keys works too. Add an Author record type, declare %rec: Book with %type: AuthorId rec Author, and recsel resolves the join:

recsel -t Book -j AuthorId -p "Title,AuthorId_Name" books.rec

Templating goes the other way with recfmt, and the toolkit ships rec2csv, csv2rec, plus an Emacs mode if that's your religion.

Why this beats the obvious alternatives:

• Vs SQLite: the file is text. git diff shows what changed, merges work, you don't need a binary tool to read it on a server. Schema lives next to the data, not in a separate migration.
• Vs CSV + xsv/Miller: CSV has no real type system, no constraints, no keys, no multi-table joins. Add 50 columns and a few enums and CSV becomes a horror movie. recutils stays sane.
• Vs JSON/YAML + jq: jq is a query language for shape; recutils gives you a query language for data — predicates, aggregates, joins, group-by — without writing reduce expressions.
• Vs a "real" database: no server, no schema migrations, no ORM. Perfect for personal datasets, lab notebooks, build manifests, asset inventories, package metadata, that one config file with 200 entries.

Install with apt install recutils or brew install recutils. The manual is one of the better GNU info documents — short, complete, full of examples. If you've ever started a side project with data.json and watched it metastasize, this is the tool you wish you'd reached for.