rowback

Surgically restore deleted rows from a massive MySQL/MariaDB dump.

You ran a DELETE you regret, and all you have is last night's multi‑gigabyte mysqldump. rowback reverses that DELETE as a filter: paste the query, and it extracts exactly the rows it would have removed — plus their foreign‑key‑linked children — and emits replayable restore SQL. No staging database, no full restore.

• Streams the dump — never loads it into memory, so a 30 GB file is fine.
• One‑time table index — the first run records each table's byte range (cached in ~/.rowback/indexes); every later query seeks straight to the tables it needs instead of re‑reading the whole file.
• Follows foreign keys — optionally pulls every child row that references the matched rows (transitively), emitted parents‑first so the restore loads cleanly.
• Two UIs — a guided web UI and a native desktop app, plus a scriptable CLI.
• Preview & curate — inspect the matched rows in a table grid, filter them, remove rows, bulk‑ or single‑cell‑edit a column, then export the curated set.

---

How it works

A mysqldump stores each table's data in one contiguous block, in alphabetical order, with no row‑level index. rowback:

1. Indexes the dump once — a fast marker scan that records every table's [start,end) byte range, saved next to nothing you care about (~/.rowback/indexes/<dump>-<hash>.idx.json).
2. Parses your query (a DELETE/SELECT/UPDATE). The WHERE clause is the filter. Supported predicates: col IN (…) literal lists, col = v, col != v, and one col IN (SELECT c FROM t WHERE …) sub‑query, combined with AND.
3. Resolves sub‑queries by reading only the referenced table's region.
4. Scans the target table's region and emits matching rows, with a quote/ escape/paren‑aware tuple parser that keeps JSON columns intact.
5. (Optional) follows foreign keys read from the dump's CREATE TABLE constraints, scanning each involved child table once (topological order), emitting their rows parents‑first.

Install / build

Requires Go 1.22+.

go build -o rowback ./cmd/rowback

Usage

Web UI (recommended)

./rowback -serve
# → http://127.0.0.1:8765

Pick the dump, paste a query, choose the output format, optionally enable related entities, run it, then curate the results in the preview grid and Export kept rows.

CLI

./rowback \
  -dump dump.sql \
  -query "DELETE FROM orders
          WHERE customer_id IN (SELECT id FROM customers WHERE region_id = 42)
            AND order_no IN ('A-1001','A-1002')" \
  -related \
  -format insert \
  -insert-mode insert \
  -out restore.sql

Flag	Default	Description
-dump	dump.sql	Path to the mysqldump .sql file
-query	(example)	DELETE/SELECT/UPDATE whose WHERE selects rows
-out	restore_reservations.sql	Output SQL file
-format	insert	insert (replayable) or raw (bare tuples)
-insert-mode	insert	insert · ignore · replace
-batch	0	Rows per INSERT statement (0/1 = one per row)
-related	false	Also extract child rows via FK constraints (transitive)
-index-dir	~/.rowback/indexes	Where the cached index lives
-serve	false	Start the web UI instead of a one‑shot run
-progress	human	human or json (one Progress object per line)

Desktop app (macOS)

cd desktop && ./build-app.sh
open "Rowback.app"

The desktop app wraps the same web UI. By default it builds a pure‑Go launcher that opens the UI in your browser; with a working C++ toolchain it builds a true native WebView window (go build -tags webview). See desktop/README.md.

Output

• insert → INSERT INTO \t` VALUES (…);(orINSERT IGNORE/REPLACE), wrapped in SET FOREIGN_KEY_CHECKS=0/…` so it loads regardless of FK order.
• raw → bare (…) tuples, one per line, for inspection or piping.

Restore by replaying the file:

mysql -h <host> -u <user> -p <db> < restore.sql

Coverage report

For each IN (…) list, rowback reports how many of the requested values were actually found — e.g. order_no 2/2, or 1/2 — missing: A-1002 — so you know whether the dump really contained everything you asked for.

Safety notes

• rowback only reads the dump and writes an SQL file; it never touches a live database.
• A WHERE clause is required — it refuses to match every row.
• Related extraction follows the dump's declared FK constraints; references to a non‑primary‑key column are noted and skipped.

License

MIT © Alberto Vincenzi