sverify

Sverify proves a data pipeline's contracts before it runs, takes the proof to DuckDB and Polars by a route that says what carries over, also runs it on DataFusion and Spark, and keeps the result up to date from Delta Lake and Iceberg change feeds.

In sverify you write a data transformation as usual, in Ibis or SQL, and state what it assumes about its inputs (requires) and what it guarantees about its output (ensures). A checker written in Lean proves the guarantees for every possible input, or says why it can't (what is proved). Then:

• Contracts are proved before the job runs. not_null, unique, ranges and relations between columns hold on every input that meets the preconditions, and division by zero, overflow and bad casts are ruled out before the job runs. The preconditions you state, and those Sverify infers, are checked on each batch before the run, or proved from Delta and Iceberg metadata, and a batch that breaks one is refused before anything runs. The runtime-error theorems list what is ruled out.
• A route takes the proof to the engine. route="auto" runs the plan on DuckDB or Polars natively, through the reference's proved translation, or by a check that the engine's choices are within the proof's, and names the plan's class there: exact, equal under an observation the postcondition states (a float tolerance is one), or exact or observed after a proved shim. Where no route applies, the plan is refused with the reason (D13). DuckDB and Polars are characterized fact by fact, with the evidence for each (the ledgers).
• Postconditions say what the result's consumer observes. By default, doubles within the plan's proved bound (floats="plan", which the report prints); sv.equivalent(floats=sv.within(abs=1e-6)) states a tolerance, which verify checks against that bound, and also rows as a sequence or a multiset, one zero. A sink states what a rerun must reproduce (reproducible=), and a refresh whose engine can't promise it is refused.
• Results are kept up to date from the changes, by delta plans proved equal to a full recompute (Rel.delta_correct), read from Delta's Change Data Feed or Iceberg's appends, and written to sinks exactly once, with the contract attached.

Eight examples, from a first contract to Substrait's own semantics, show each of these in a few lines.

We aim to make sverify an assistant that makes data engineering more joyful: fewer 3 a.m. pages, fewer surprises between engines, more trust in the tables you hand on.

Sverify is purely observational. It's a tool to explore and compare database semantics. Although it is already quite useful, production shouldn't depend on it yet, as the API will change (sometimes drastically) between minor releases. It describes each engine as observed at pinned versions, without certifying or ranking it, and isn't affiliated with any engine's project or vendor.

Sverify checks every plan against a reference semantics, and four principles decide what it does (reference-semantics.md):

1. The reference follows Substrait, options included, and says where it departs from it.
2. Where Substrait is silent, it takes the least surprising choice and writes down why. An option a plan omits means an error.
3. Every engine can follow it with reasonable effort.
4. Optimizations stay provable: an error means "may raise".

The reference raises where engines disagree on an answer, so a verified plan runs alike on each: an operation on finite doubles that overflows to ±∞, power(-inf, 0.5), a MEDIAN between -inf and +inf, a constant subexpression that raises. AND and OR evaluate both operands, as DuckDB and Polars do. A plan verified under the reference creates no NaN and no ±∞ its inputs did not hold. semantics="duckdb" checks a plan under what DuckDB does, and semantics="substrait" under Substrait's extension YAMLs, open wherever they say nothing (the Substrait profile).

Engines come in three tiers (#211):

Work on DataFusion, Spark and Snowflake is paused until the generated lowerings land (#241, then #271). DataFusion is not installed by default (the datafusion extra adds it).

Polars refuses what only DuckDB runs so far, with "see #271": the functions that SUPPORT and WINDOW_SUPPORT list without polars (regexp_extract, regexp_replace, md5, sha256, first, last, string_agg, median, the quantiles, mode, arg_min, arg_max, the covariances, corr, and variances over a window), the black boxes, and MIN and MAX of doubles (#697).

pip install sverify                # or: uv add sverify
pip install 'sverify[all]'         # with every engine and table format

The core runs on DuckDB and writes DuckDB, Snowflake and Spark SQL. Extras: datafusion, polars, spark (needs Java 17 or 21), delta, iceberg, flight, testing, or all. Wheels bundle the checker as a small native binary for Linux (x86_64, aarch64) and macOS (arm64), Python 3.12 and later; elsewhere pip builds from the sdist, which needs elan and starts a full lake build of every generated Lean module, minutes of work that can load every core. Contributors with a built checker set SVERIFY_PREBUILT_CHECK=<path to substrait-check>, and an editable install links it instead.

Each night, when main has changed, the nightly pre-release gets its wheels, versioned <next minor>.dev<commits>+g<sha>, so they sort after the latest release. They are not on PyPI:

uv add "sverify>=0.4.0.dev0" --find-links https://github.com/AlexanderNenninger/substrait-lean/releases/expanded_assets/nightly
pip install --pre sverify --find-links https://github.com/AlexanderNenninger/substrait-lean/releases/expanded_assets/nightly

uv keeps the version and drops the page, so add it to your pyproject.toml; uv lock --upgrade-package sverify then moves to the latest nightly:

[tool.uv]
find-links = ["https://github.com/AlexanderNenninger/substrait-lean/releases/expanded_assets/nightly"]

Each build's wheels stay on the release for 30 days, so a wheel's own URL pins one build for that long. --find-links sees every build on the page and picks the newest. Copy a URL from the release page:

pip install "sverify @ https://github.com/AlexanderNenninger/substrait-lean/releases/download/nightly/sverify-<version>-py3-none-<platform>.whl"

The API reference documents every public function and class of the released version, with the examples and the ledgers: the upstream bugs and Substrait spec gaps sverify found, and each engine's characterization, searchable by id, engine, function and option. For coding agents, the site also publishes skills for writing contracts, pipelines and engine code with sverify (npx skills add https://alexandernenninger.github.io/substrait-lean/); they ship in the wheel too, under sverify/skills/.

In thirty lines

import pyarrow as pa
import sverify as sv
from ibis import _
from sverify.io import Changes, StaticSource

orders = StaticSource(  # in memory; DeltaSource and IcebergSource work the same way
    pa.table({"id": [1, 2], "customer": ["ada", "bob"], "amount": [30, 5]}), name="orders"
)


@sv.transform(
    requires={"orders": [_.amount >= 0, _.amount < 1_000_000_000, sv.unique("id")]},
    ensures=[_.doubled >= 22, _.doubled.notnull(), sv.unique("id")],
)
def big_orders(orders):
    return orders.filter(orders.amount > 10).select(orders.id, doubled=orders.amount * 2)


big = orders.pipe(big_orders)  # proved here, before any data runs
big.refresh()  # computed, on DuckDB by default
orders.apply(Changes(inserts=[(3, "cy", 50)], deletes=[(1, "ada", 30)]))
big.refresh(engine="polars", route="auto")  # updated from the changes alone, on Polars
assert big.to_arrow().to_pylist() == [{"id": 3, "doubled": 100}]
print(big_orders.report)
# requires orders: (_.amount >= 0)
# requires orders: (_.amount < 1000000000)
# requires orders: unique(id)
# ensures (_.doubled >= 22): proved

A postcondition that doesn't follow is refused before anything runs:

@sv.transform(requires={"orders": [_.amount >= 0]}, ensures=[_.total > 5])
def plus_one(orders):
    return orders.select(orders.id, total=orders.amount + 1)


from sverify.errors import VerificationError

try:
    plus_one(orders)
except VerificationError as e:
    print(e)
# 1 of 1 postcondition not proved, 1 precondition needed (e.report has the details)
#   ...
#   not proved: (_.total > 5): it may be FALSE on some rows ...
#     would hold with `(_.amount >= 5)` in the requires of 'orders'

The examples go on from here: SQL, inferred preconditions, routes and their classes, tolerances for doubles, reproducible sinks, engine profiles and Substrait's own semantics.

Testing with pytest

sverify.testing has assertions for your test suite, and sverify registers a pytest plugin with two fixtures: sv_source makes an in-memory source from rows, and a test that takes sv_engine runs once per installed engine.

from sverify.testing import assert_contract, assert_engines_agree, source


def test_big_orders():
    orders = source([], {"id": "int64", "customer": "string", "amount": "int64"}, name="orders")
    node = orders.pipe(big_orders)
    assert_contract(node)  # the contract is proved
    data = {"orders": [(1, "ada", 30), (2, "bob", 5)]}
    assert_contract(node, data=data, expected=[(1, 60)])  # and holds on these rows
    assert_engines_agree(node, data)  # every installed engine agrees


test_big_orders()

A failure lists the offending rows: those that break a precondition, or the expected rows that are missing or unexpected. sverify.testing.cases(node) is a Hypothesis strategy of inputs that meet the preconditions, for property tests of the code around a node (pip install 'sverify[testing]').

SQL models

A transform's body can be SQL over its inputs by name. Sverify proves, composes and refreshes it like any other:

@sv.transform(
    requires={"orders": ["amount >= 0", "amount IS NOT NULL"]},
    ensures=["total >= 0", sv.unique("customer")],
)
def per_customer(orders):
    return sv.sql(
        "SELECT customer, coalesce(sum(amount), 0) AS total FROM orders GROUP BY customer",
        orders=orders,
    )


totals = orders.pipe(per_customer)  # proved here, like an Ibis transform
totals.refresh()
assert sorted(totals.to_arrow().to_pylist(), key=str) == [
    {"customer": "bob", "total": 5},
    {"customer": "cy", "total": 50},
]

Several inputs are several keywords (sv.sql(query, orders=orders, customers=customers)). The SQL is DuckDB's, and sverify plans it itself, through SQLGlot, with sverify's reference semantics (semantics="reference", the default) or DuckDB's (semantics="duckdb"), for a subset of SQL so far (#239, #240).

Workflows

Three complete jobs from a catalogue of 139 real analytical queries, each from source tables to a sink, with what goes wrong without the checker:

• Duplicate card payments: a self-join over a Delta table, refreshed from its Change Data Feed, where a watermark job would miss a late duplicate and keep a reversed one.
• Sensor alerts: LAG over an Iceberg table, where ties and a NaN reading change the alerts, and Snowflake and Spark need proved tie-freedom.
• An orders mart: dbt tests as a proved contract on a point-in-time join to a type-2 dimension, in Ibis and in SQL.

The whole tour adds two guides, to contracts and to pipelines and engines; it runs locally with uvx marimo edit --sandbox notebooks from a clone.

The API in brief

import ibis
from sverify.engines import SNOWFLAKE, to_sql  # also DUCKDB, SPARK
from sverify.io import DeltaSink, DeltaSource  # IcebergSource, IcebergSink likewise

sv.verify(
    "select id, amount from orders where amount > 0",
    tables={"orders": ibis.table({"id": "int64", "amount": "int64"}, name="orders")},
    ensures=["amount > 0"],
)  # SQL, checked; returns the report
to_sql(big.internal_plan, SNOWFLAKE)  # the checked plan, or sverify.engines.Unsupported

# p = DeltaSource("s3://.../orders").pipe(paid).write(DeltaSink("s3://.../paid"))  # stored
# r = p.pipe(revenue, regions=regions).write(DeltaSink("s3://.../revenue"))
# r.refresh(engine="polars"); sv.group(r, other).refresh(); r.to_pandas()

• @sv.transform(requires=..., ensures=..., on_unproved="error" | "runtime" | "warn") proves a contract on each call; .report says what was proved and what is known about every output column. sv.unique(...) states keys.
• source.pipe(paid).pipe(revenue, regions=regions) builds nodes left to right, in the order the data flows (the call form revenue(paid(source), regions=regions) is the same node). A step is a transform, SQL (sv.sql(query), which binds the chained node to the one table the keywords leave over) or a plain function. A downstream precondition is proved from the upstream contract (CompositionError names a broken link).
• Without .write in between, transforms fuse into one plan; .write(sink) stores a node, exactly once per state, and nodes downstream read it as a versioned source. refresh(engine=...) updates the whole lineage, and sv.group(a, b) several outputs against the same source versions.
• Filters, projections, inner and cross joins and UNION ALL are maintained from the changes. An aggregate recomputes only the groups a batch touches, and a window only the partitions it touches; anything else is recomputed and diffed.
• sverify.io has Delta, Iceberg and in-memory sources and sinks, and base classes for your own: a Source implements read(since) (from_query, from_frame and from_files build one), a Sink implements write(change) (sink_from wraps a function).
• sverify.testing has assert_contract, assert_engines_agree and the Hypothesis strategy cases for pytest, with the plugin's sv_engine and sv_source fixtures.
• report.run(con, engine, route="auto") runs a verified plan on an engine by a route (report.routes records it, with its class), and refresh(engine=..., route="auto") does the same for each node. sv.equivalent(...) in ensures names the observation, and reproducible= on a sink what a rerun must reproduce.
• sverify.engines has the engines, SQL dialects and plans (run_sharded runs a plan per shard and combines the results, with a proof that the result is the same); sverify.errors the exceptions.

Not supported yet: correlated scalar subqueries, unbound parameters (sv.verify(expr, params={p: 10}) checks an ibis.param with its value), sample, random() and now(), among the operations the Ibis coverage table marks ✗. The repository has a guide, examples, the Ibis coverage table in its README, and what is proved and what is trusted.

Configuration

• SVERIFY_CHECK: path to a substrait-check binary that replaces the bundled one, for example your own lake build.
• DataFusion is pinned to 51, because Ibis 12's DataFusion backend doesn't connect to later versions.

Development

git clone https://github.com/AlexanderNenninger/substrait-lean
cd substrait-lean/python
uv sync                          # builds substrait-check with lake (needs elan) and links it in
uv run poe qa                    # proofs, lint, typecheck, golden, Python tests, notebooks

The repository has setup details.

License

Sverify is licensed under the Mozilla Public License 2.0 (MPL-2.0). You can use it in any product, open or proprietary, and your own code stays yours. If you distribute a modified version of sverify's own files, those files stay under the MPL and you share their source. The license applies file by file, so it never reaches the code around them.

python/src/sverify/ibis_substrait/ is derived from ibis-substrait and stays under the Apache License 2.0.