Documentation · Live demo · Case study · Quickstart · Test results

Rust WebAssembly MIT

Fishbowl is a SQL database written from scratch in Rust. It passes 100% of SQLite's 5.9 million sqllogictest records, and it runs in your browser, where you can watch each query run row by row.

Watch the 30-second demo

Type a query and each row rises out of its table, through every step of the plan, into the result. You can see why each step was chosen, which rows an index let the query skip, and the B-tree pages the data lives on. The live demo is the real engine compiled to WebAssembly, with no server.

Correctness, measured

Fishbowl runs SQLite's own sqllogictest suite, at check-in ab599cbdd0976a63c5cfc7755588f716a7621f27 (2026-10-07).

Corpus	Files	Records run	Failed	Skipped	Pass rate	Queries only
Full suite	622	5,939,879	0	1,480,834	100.00%	100.00%
29-file sample	29	129,348	0	14,153	100.00%	100.00%

The full suite is every test file in the repository at that check-in, and gives the headline number. The sample is the fast check CI runs on every push. Skipped records are ones a file marks as not applying to SQLite, such as onlyif mysql; they are never counted as passes. Every file is pinned by SHA-256, "not supported yet" counts as a failure, and the report is byte-identical from run to run. Passing does not mean all of SQL works: the suite never uses window functions, common table expressions or ALTER TABLE. RESULTS.md lists the last failures and how they were fixed.

The suite is one of five checks. A differential test runs random scripts through Fishbowl and the real sqlite3 and requires the same answers. Property tests compare the B-tree with Rust's BTreeMap. Bugs planted on purpose check that the tests notice. A crash test cuts the power at each of the 178 points where a workload writes to or syncs the disk, and checks after 7,189 recoveries that every committed transaction is there and nothing else is. The case study covers all five.

What works

Queries can join tables, nest subqueries (correlated ones too), combine results with UNION, INTERSECT and EXCEPT, and read views, along with sorting, grouping and aggregates. INSERT, UPDATE and DELETE write to B-trees on 4 KiB pages. PRIMARY KEY, UNIQUE and CREATE INDEX are enforced on every write. A cost-based planner, using SQLite's cost model and statistics from ANALYZE, chooses between a full scan and each index, orders joins, looks up matching rows by index inside a join, and lets an index stand in for a sort. It keeps the plans it turned down, with their costs, for the UI. BEGIN, COMMIT and ROLLBACK group changes into transactions, which commit through a write-ahead log. Triggers fire before, after or instead of each changed row. Anything else is rejected with "not supported yet" rather than answered wrongly.

Part	State
Parse, plan, execute pipeline	Joins, subqueries, compound SELECTs, views, sorting, grouping, aggregates
Storage	Pager, buffer pool and B+tree tables on 4 KiB pages (M3)
Indexes and constraints	PRIMARY KEY, UNIQUE, CREATE INDEX (M4)
Cost-based planner	ANALYZE, SQLite's cost model, composite and covering indexes, index lookups in joins, join order search, rejected plans shown (M5)
Transactions and crash recovery	BEGIN/COMMIT/ROLLBACK, write-ahead log, checkpoints, recovery on open, crash test (M6)
Browser UI	Row-by-row trace playback, why this plan with the plans turned down, B-tree page viewer with live splits (M7, M5)
Proof	Full sqllogictest suite, CI workflow, case study (M8)

Run it

You need Rust (via rustup) with the wasm32-unknown-unknown target, wasm-bindgen-cli 0.2.129, Node.js, pnpm, and the sqlite3 CLI for the differential test. Getting started has more detail.

# Engine tests
cargo test --workspace

# SQL shell
cargo run -p fishbowl-cli -- -c "
  CREATE TABLE fish (id INTEGER PRIMARY KEY, name TEXT, species TEXT);
  INSERT INTO fish VALUES (1, 'goldie', 'goldfish'), (2, 'nemo', 'clownfish'),
                          (3, 'koko', 'koi'), (4, 'pearl', 'koi');
  CREATE INDEX fish_species ON fish (species);
  ANALYZE;
  SELECT name FROM fish WHERE species = 'koi';"

name
-----
koko
pearl
(2 rows)

plan:
Project name  [rows=2]
  Filter (species = 'koi')  [rows=2]
    SeqScan fish  [rows=4]
      why: reading all 4 rows of fish is the cheapest way here
      estimate: rows=4 cost=12
      rejected: IndexScan fish USING INDEX fish_species (species = 'koi')  [rows=2 cost=13]

On four rows, reading them all is cheaper than finding two through the index. On a big table the index wins.

# Browser UI at http://localhost:5173
cargo install wasm-bindgen-cli --version 0.2.129
pnpm --dir web install && pnpm --dir web dev

# Crash test: prints how many crash points it tried
cargo test -p fishbowl --test crash -- --nocapture

# sqllogictest, the 29-file sample: 26 MB, about 2 seconds
scripts/fetch-sqllogictest.sh
cargo run --release -p fishbowl-slt -- --write

# sqllogictest, the full suite: 1.2 GB, about 40 seconds on 10 cores
scripts/fetch-sqllogictest.sh --full
cargo run --release -p fishbowl-slt -- --full --write

--write updates the committed reports in sqllogictest/. The corpus downloads into sqllogictest/test, a symlink to test.nosync so iCloud Drive skips it.

Layout

Path	What
crates/engine	The engine: values, catalog, planner, executor, storage
crates/cli	fishbowl SQL shell
crates/slt	sqllogictest runner and report
crates/wasm	WebAssembly bindings
web/	React, Vite and Tailwind UI. A static site with no server
website/	The docs site, a Fumadocs app that renders docs/
sqllogictest/	Corpus manifests, results and generated reports
scripts/	Fetch and pin the sqllogictest corpus
docs/	The docs: case study, architecture, milestones, harness notes, learning notes
brand/	Logo, colors and type. See BRAND.md

License

MIT. The sqllogictest files are downloaded from sqlite.org and are not part of this repository; see their terms.