cygnet-orm 1.1.0

The littlest ORM — bring your own objects, write real SQL

CYGNET

A small but fierce ORM. Bring your own objects, write real SQL.

CYGNET is a PostgreSQL-first ORM for Python 3.12+ that keeps SQL visible. SQL keywords are uppercase method names. Python utilities are lowercase. If you're comfortable with SQL and just want some help writing it, CYGNET is for you.

It's small but full-featured: dataclass models, foreign keys with FOLLOW/LEFT_FOLLOW joins, upsert via ON CONFLICT DO UPDATE, LATERAL joins and EXISTS / IN subqueries, FOR UPDATE row locking, savepoint-aware transactions, full mypy strictness, and an expression protocol that lets cygnet.op() / cygnet.ops() / cygnet.lit() extend the query API without touching internals.

Documentation

This README covers installation and usage. For the internals:

• ARCHITECTURE.md — module map, load-bearing invariants, and landmines. The fast orientation for picking up the code cold.
• THEORY.md — the theory of operation: why Cygnet is built the way it is, the rejected alternatives, and the tradeoffs.
• GitHub Issues — open issues and bug reports.

Installation

pip install cygnet-orm                # core only — bring your own db adapter
pip install 'cygnet-orm[psycopg]'     # + reference psycopg3 adapter
pip install 'cygnet-orm[asyncpg]'     # + reference asyncpg adapter

Cygnet itself doesn't depend on a particular database driver: the db object is duck-typed (four methods — see The db object below). If you want the bundled PsycopgDB reference adapter, install the [psycopg] extra; otherwise the core install stays driver-free and you supply your own execute / execute_one / stream methods.

Requires Python 3.12+ and PostgreSQL 14+.

Quick start

import dataclasses
from typing import Annotated
import cygnet

@dataclasses.dataclass
class Account:
    id: Annotated[int, cygnet.DBKey]   # database-assigned primary key
    name: str
    email: str

AccountTable = cygnet.Table(Account)

SELECT

# All rows
accounts = await cygnet.SELECT(db).FROM(AccountTable)

# With WHERE
accounts = await cygnet.SELECT(db).FROM(AccountTable).WHERE(
    AccountTable.name == "Fred"
)

# Compound predicates
accounts = await cygnet.SELECT(db).FROM(AccountTable).WHERE(
    (AccountTable.name == "Fred") & (AccountTable.id > 10)
)

# Specific columns — returns list of tuples
rows = await cygnet.SELECT(db, AccountTable.id, AccountTable.name).FROM(AccountTable)

# Pagination — ORDER BY + LIMIT + OFFSET
accounts = await (
    cygnet.SELECT(db)
    .FROM(AccountTable)
    .ORDER_BY(AccountTable.name)
    .LIMIT(20)
    .OFFSET(40)        # third page of 20
)

# ORDER BY descending
accounts = await (
    cygnet.SELECT(db)
    .FROM(AccountTable)
    .ORDER_BY(AccountTable.name, DESC=True)
)

# Mixed ASC/DESC: chain multiple ORDER_BY calls.  A single call applies
# one direction to all of its arguments.
accounts = await (
    cygnet.SELECT(db)
    .FROM(AccountTable)
    .ORDER_BY(AccountTable.name)                  # ASC
    .ORDER_BY(AccountTable.id, DESC=True)         # DESC
)

# GROUP BY (requires explicit columns)
rows = await (
    cygnet.SELECT(db, AccountTable.name)
    .FROM(AccountTable)
    .GROUP_BY(AccountTable.name)
)

JOIN

@dataclasses.dataclass
@cygnet.table("log_entries")       # override table name
class LogEntry:
    id: Annotated[int, cygnet.DBKey]
    account_id: int
    message: str

LogTable = cygnet.Table(LogEntry)

# INNER JOIN — returns list of (Account, LogEntry) tuples
rows = await (
    cygnet.SELECT(db)
    .FROM(AccountTable)
    .JOIN(LogTable, ON=AccountTable.id == LogTable.account_id)
)

# LEFT JOIN — right side is None when there is no match
rows = await (
    cygnet.SELECT(db)
    .FROM(AccountTable)
    .LEFT_JOIN(LogTable, ON=AccountTable.id == LogTable.account_id)
)
for account, entry in rows:
    print(account.name, entry.message if entry else "(no log)")

# RIGHT JOIN — left (FROM-side) is None when there is no match
rows = await (
    cygnet.SELECT(db)
    .FROM(AccountTable)
    .RIGHT_JOIN(LogTable, ON=AccountTable.id == LogTable.account_id)
)
for account, entry in rows:
    print(account.name if account else "(orphan log)", entry.message)

# FULL JOIN — either side can be None (matched rows populate both)
rows = await (
    cygnet.SELECT(db)
    .FROM(AccountTable)
    .FULL_JOIN(LogTable, ON=AccountTable.id == LogTable.account_id)
)
for account, entry in rows:
    a = account.name if account else "(no account)"
    e = entry.message if entry else "(no log)"
    print(a, e)

# Self-join via aliases — same table referenced twice in one query
A = AccountTable.AS("a")
B = AccountTable.AS("b")
pairs = await (
    cygnet.SELECT(db, A.name, B.name)
    .FROM(A)
    .JOIN(B, ON=A.id != B.id)
)

Lateral joins

LATERAL subqueries reference columns from preceding FROM / JOIN tables — the canonical "top-N per group" pattern, or any per-row correlated subquery a regular JOIN can't express:

# Most recent log entry per account
recent = (
    cygnet.SELECT(db, LogTable.message)
    .FROM(LogTable)
    .WHERE(LogTable.account_id == AccountTable.id)   # outer reference
    .ORDER_BY(LogTable.id, DESC=True)
    .LIMIT(1)
)
recent_lat = cygnet.lateral("recent", recent, columns=["message"])

rows = await (
    cygnet.SELECT(db, AccountTable.name, recent_lat.message)
    .FROM(AccountTable)
    .LEFT_JOIN_LATERAL(recent_lat)        # NULL for accounts with no logs
)

JOIN_LATERAL and LEFT_JOIN_LATERAL accept an optional ON= predicate (defaults to cygnet.lit("true") since PG syntax requires ON even when there's no extra filter to express). Column inference follows the same rules as cygnet.cte(): explicit ColumnProxy projections work without columns=[…]; opaque expressions need it.

INSERT

# From a dataclass instance — DBKey fields with None are omitted,
# and the generated key is written back onto the object via RETURNING
acc = Account(id=None, name="Fred", email="[e-mail]")
await cygnet.INSERT(db).INTO(AccountTable).VALUES(acc)
print(acc.id)   # populated by PostgreSQL

# From keyword arguments
await cygnet.INSERT(db).INTO(AccountTable).VALUES(
    name="Wilma", email="[e-mail]"
)

# Bulk INSERT — one statement, many rows, one round-trip.  Each
# object's DBKey is populated in input order from the RETURNING result.
accs = [
    Account(id=None, name="Fred", email="[e-mail]"),
    Account(id=None, name="Wilma", email="[e-mail]"),
    Account(id=None, name="Barney", email="[e-mail]"),
]
ids = await cygnet.INSERT(db).INTO(AccountTable).BULK_VALUES(accs)
# ids == [acc.id for acc in accs]

# INSERT … SELECT — copy or transform rows in one statement.  Target
# columns are inferred from the source's ColumnProxy projection;
# pass columns=[…] explicitly when the source uses opaque expressions.
source = (
    cygnet.SELECT(db, AccountTable.name, AccountTable.email)
    .FROM(AccountTable)
    .WHERE(AccountTable.created_at > cutoff)
)
new_ids = await cygnet.INSERT(db).INTO(ArchiveTable).SELECT(source)

UPDATE

# Partial update via kwargs
await (
    cygnet.UPDATE(db)
    .SET(AccountTable, name="Frederick")
    .WHERE(AccountTable.id == 1)
)

# Full object update (pk excluded from SET clause)
acc.name = "Frederick"
await cygnet.UPDATE(db).SET(AccountTable, acc).WHERE(AccountTable.id == acc.id)

# Expressions on the right-hand side render in place rather than as
# parameters — that's how `count = count + 1` and any computed update
# work.  Any SQLRenderable (column ref, op(), fn(), lit(), …) is fine.
await (
    cygnet.UPDATE(db)
    .SET(AccountTable, name=cygnet.fn("upper")(AccountTable.name))
    .WHERE(AccountTable.id == 1)
)

# Cross-table UPDATE: pull values from another table.  PG joins via the
# WHERE clause (not a separate JOIN/ON), unlike SELECT.
await (
    cygnet.UPDATE(db)
    .SET(AccountTable, email=LogTable.message)
    .FROM(LogTable)
    .WHERE(AccountTable.id == LogTable.account_id)
)

# RETURNING gives back the updated rows (a list of tuples).
[(new_name,)] = await (
    cygnet.UPDATE(db)
    .SET(AccountTable, name="Updated")
    .WHERE(AccountTable.id == 1)
    .RETURNING(AccountTable.name)
)

UPDATE requires an explicit .WHERE() clause — a mass-mutation safety rail. Pass cygnet.all to opt in: .WHERE(cygnet.all).

DELETE

# DELETE always requires WHERE (the same safety rail).
await cygnet.DELETE(db).FROM(LogTable).WHERE(LogTable.account_id == 42)

# To wipe every row, opt in with cygnet.all.
await cygnet.DELETE(db).FROM(LogTable).WHERE(cygnet.all)

# Cross-table DELETE: USING references other tables; the join condition
# lives in WHERE.  PG's syntactic mirror of UPDATE … FROM.
await (
    cygnet.DELETE(db)
    .FROM(LogTable)
    .USING(AccountTable)
    .WHERE(
        (LogTable.account_id == AccountTable.id)
        & (AccountTable.name == "Fred")
    )
)

# RETURNING gives back the deleted rows.
deleted = await (
    cygnet.DELETE(db)
    .FROM(LogTable)
    .WHERE(LogTable.account_id == 42)
    .RETURNING(LogTable.id, LogTable.message)
)

For "drop every row" workflows, cygnet.TRUNCATE is faster (acquires a stronger lock and resets sequences):

await cygnet.TRUNCATE(db, LogTable)
await cygnet.TRUNCATE(db, LogTable, AccountTable, cascade=True)

ON CONFLICT — explicit conflict handling

For finer control than save(), the INSERT builder exposes PG's full ON CONFLICT family. Currently scoped to single-row VALUES(obj); bulk + INSERT…SELECT variants will land in a follow-up.

# Skip the row if it conflicts with any unique constraint.
# Returns None on skip; the object's PK is left unset.
await (
    cygnet.INSERT(db).INTO(AccountTable).VALUES(acc)
    .ON_CONFLICT_DO_NOTHING()
)

# Skip if conflict on a specific column / set of columns.
await (
    cygnet.INSERT(db).INTO(AccountTable).VALUES(acc)
    .ON_CONFLICT(AccountTable.email).DO_NOTHING()
)

# Or skip if a named constraint fires.
await (
    cygnet.INSERT(db).INTO(AccountTable).VALUES(acc)
    .ON_CONFLICT_CONSTRAINT("uq_accounts_email").DO_NOTHING()
)

# DO UPDATE: rewrite the existing row with literal kwarg values.
await (
    cygnet.INSERT(db).INTO(AccountTable).VALUES(acc)
    .ON_CONFLICT(AccountTable.email)
    .DO_UPDATE(name="Updated", email="[e-mail]")
)

# DO UPDATE FROM EXCLUDED: rewrite the existing row with the values
# the new row tried to insert.  This is what save() does internally.
await (
    cygnet.INSERT(db).INTO(AccountTable).VALUES(acc)
    .ON_CONFLICT(AccountTable.email)
    .DO_UPDATE_FROM_EXCLUDED(AccountTable.name, AccountTable.email)
)

When DO NOTHING skips a row, INSERT … RETURNING returns no rows and the awaited builder yields None. This is the only case where an empty RETURNING is treated as a normal outcome — without ON_CONFLICT, an empty RETURNING still raises (the same silent-failure guard as before).

save() — upsert

# New object (DBKey + None) -> INSERT ... RETURNING, pk populated
acc = Account(id=None, name="Fred", email="[e-mail]")
await cygnet.save(db, acc)

# Existing object -> INSERT ... ON CONFLICT DO UPDATE
acc.name = "Frederick"
await cygnet.save(db, acc)

save() is DEFAULT-aware (matches INSERT and create() since 2026-05-22): a field whose in-memory value is None and whose column has a schema DEFAULT is omitted from both the INSERT column list and the DO UPDATE SET clause, then refreshed via RETURNING. Consequences:

• New row (no conflict): the schema DEFAULT fires (e.g. now() for created_at), and obj.created_at is patched with the populated value.
• Existing row (conflict): the DEFAULTed column is not touched by the UPDATE — the existing value is preserved — and RETURNING still refreshes obj to match the DB row.
• Explicit override: a non-None value is always written through, so the app can override the DEFAULT when it wants to.

In other words, obj.created_at = None is now a signal to "leave the DB's value alone"; use UPDATE if you need to write a literal NULL to a DEFAULTed column. Adapters that don't implement the optional column_defaults protocol method (e.g. FakeDB) see the historical shape: every field emitted, no RETURNING.

For surgical updates of individual columns, use UPDATE.

get() — fetch by primary key

acc = await cygnet.get(db, AccountTable, id=42)  # Account | None

create() — INSERT without ON CONFLICT

# Equivalent to INSERT … RETURNING; duplicate-key violations propagate
# from the database rather than being silently upserted.  Use this when
# you want the database to tell you "this already existed" loudly.
acc = await cygnet.create(db, Account(id=None, name="Fred", email="[e-mail]"))

Foreign keys and FOLLOW

Annotate a field with cygnet.ForeignKey(Target) to declare a foreign key. Cygnet validates the reference at introspection time (target must be a dataclass with a PK; types must match) and uses it to power FOLLOW / LEFT_FOLLOW and cygnet.follow:

@dataclasses.dataclass
@cygnet.table("log_entries")
class LogEntry:
    id: Annotated[int, cygnet.DBKey]
    account_id: Annotated[int, cygnet.ForeignKey(Account)]
    message: str

LogTable = cygnet.Table(LogEntry)

# FOLLOW is an INNER JOIN with the FK condition spelled out for you.
# Returns tuples of (LogEntry, Account):
rows = await cygnet.SELECT(db).FROM(LogTable).FOLLOW(LogTable.account_id)

# LEFT_FOLLOW for the outer-join variant; the right side is None when
# the FK is NULL or the target row is missing.
rows = await cygnet.SELECT(db).FROM(LogTable).LEFT_FOLLOW(LogTable.account_id)

# Load a single related object on demand (returns Account | None).
log = await cygnet.get(db, LogTable, id=1)
account = await cygnet.follow(db, log, LogTable.account_id)

ForeignKey(target) always references the target's primary key — composite PKs aren't supported.

Subquery predicates: EXISTS / IN

A SelectBuilder is itself an SQL expression — it renders as a parenthesised inline subquery wherever an expression is expected. That makes EXISTS, IN (SELECT …), and scalar subqueries work without any extra builder methods:

# EXISTS — is there at least one matching row?  Correlated subqueries
# reference outer-query columns directly (no LATERAL needed).
any_log = (
    cygnet.SELECT(db, cygnet.lit("1"))
    .FROM(LogTable)
    .WHERE(LogTable.account_id == AccountTable.id)
)
authors = await (
    cygnet.SELECT(db).FROM(AccountTable).WHERE(cygnet.exists(any_log))
)

# NOT EXISTS — anti-join idiom.  ~cygnet.exists(b) works too;
# both render as `NOT EXISTS (…)` (not `NOT (EXISTS …)`).
silent = await (
    cygnet.SELECT(db).FROM(AccountTable).WHERE(cygnet.not_exists(any_log))
)

# IN (subquery) — uses cygnet.op since `in` isn't a Python overload.
active_ids = (
    cygnet.SELECT(db, AccountTable.id)
    .FROM(AccountTable)
    .WHERE(AccountTable.email == "[e-mail]")
)
fred_logs = await (
    cygnet.SELECT(db)
    .FROM(LogTable)
    .WHERE(cygnet.op(LogTable.account_id, "IN", active_ids))
)

# Scalar subquery in the SELECT list — the builder is rendered directly.
log_count = (
    cygnet.SELECT(db, cygnet.fn("count")(cygnet.lit("*")))
    .FROM(LogTable)
    .WHERE(LogTable.account_id == AccountTable.id)
)
rows = await cygnet.SELECT(db, AccountTable.name, log_count).FROM(AccountTable)

$N parameter numbering threads correctly through the inner-then-outer pieces.

Row-level locking

# FOR UPDATE — exclusive lock; blocks other locks until commit.
acc = await (
    cygnet.SELECT(db).FROM(AccountTable).WHERE(AccountTable.id == 1)
    .FOR_UPDATE()
)

# Skip rows that someone else has locked — the queue-worker pattern.
batch = await (
    cygnet.SELECT(db).FROM(LogTable)
    .ORDER_BY(LogTable.id)
    .LIMIT(10)
    .FOR_UPDATE(skip_locked=True)
)

# Restrict the lock to specific tables in a join with `of=`.
rows = await (
    cygnet.SELECT(db).FROM(AccountTable)
    .JOIN(LogTable, ON=AccountTable.id == LogTable.account_id)
    .FOR_UPDATE(of=AccountTable)            # lock accounts only
)

# FOR SHARE — shared lock; allows concurrent reads, blocks writes.
await cygnet.SELECT(db).FROM(AccountTable).WHERE(...).FOR_SHARE()

Both verbs accept nowait=True (fail immediately if any row is locked) and skip_locked=True (silently skip locked rows; mutually exclusive with nowait). The rarer modes ride as flags: FOR_UPDATE(no_key=True) → FOR NO KEY UPDATE, FOR_SHARE(key=True) → FOR KEY SHARE.

Inspecting SQL without executing

Every builder exposes .sql(), returning the rendered SQL plus the parameter list — same validation as await, but no round trip:

sql, params = (
    cygnet.SELECT(db, AccountTable.name)
    .FROM(AccountTable)
    .WHERE(AccountTable.id == 1)
    .sql()
)
# sql == "SELECT accounts.name FROM accounts WHERE (accounts.id = $1)"
# params == [1]

UPDATE / DELETE .sql() calls run the same WHERE-required validation as execution, so the safety rail isn't bypassable through inspection.

Functions and operators

import cygnet.functions as f

# Aggregates
total = await cygnet.SELECT(db, f.count(), f.sum(OrderTable.amount)).FROM(OrderTable)

# COUNT(*) is the empty-args form
rows = await cygnet.SELECT(db, f.count()).FROM(AccountTable)

# In WHERE / HAVING via comparison overloads
busy = await (
    cygnet.SELECT(db, AccountTable.name)
    .FROM(AccountTable)
    .GROUP_BY(AccountTable.name)
    .HAVING(f.count() > 1)
)

# Anything not curated is reachable via cygnet.fn(name)
await cygnet.SELECT(db, cygnet.fn("date_trunc")("day", OrderTable.created_at)).FROM(OrderTable)

For inline operators (ILIKE, ~~, @@, etc.) use cygnet.op / cygnet.ops / cygnet.is_null / cygnet.is_not_null. For raw SQL fragments, cygnet.lit("..."). Operator and function names are trusted strings — never pass user input as an operator or function name.

Comparing a column to None is NULL-safe: T.col == None renders col IS NULL and T.col != None renders col IS NOT NULL — so a value that is None at runtime does the right thing instead of silently matching no rows. cygnet.is_null(col) / cygnet.is_not_null(col) are the explicit equivalents.

cygnet.op has three arities:

# 3-arg: one-shot infix predicate
.WHERE(cygnet.op(T.name, "ILIKE", "%fred%"))

# 2-arg: prefix operator (NOT, EXISTS-style)
.WHERE(cygnet.op("NOT", T.active == True))

# 1-arg: factory — bind the operator once, reuse the callable.
# Idiomatic when the same non-standard operator appears repeatedly:
ILIKE = cygnet.op("ILIKE")
.WHERE(ILIKE(T.name, "%fred%") | ILIKE(T.email, "%fred%"))

The 1-arg form is a closure capturing the operator string and returning a (left, right) -> Predicate callable. Operands are still parameterised; only the operator string is interpolated verbatim — same trusted-string rule as the other arities.

JSONB, arrays, and full-text search

Three curated submodules wrap the most common PG-native operators and functions. Each is a thin layer over cygnet.op / cygnet.fn, so anything not curated is still reachable directly.

import cygnet.jsonb as jb
import cygnet.arrays as arr
import cygnet.fts as fts

# JSONB — `data ->> 'name' = 'Fred'`, `data @> '{"active": true}'`, etc.
.WHERE(jb.get_text(T.payload, "name") == "Fred")
.WHERE(jb.contains(T.payload, {"active": True}))
.WHERE(jb.has_key(T.payload, "email"))

# Arrays — @>, <@, &&, ANY/ALL, length
.WHERE(arr.contains(T.tags, ["python", "sql"]))
.WHERE(arr.overlaps(T.tags, ["python", "go"]))
.WHERE(T.id == arr.any(other.allowed_ids))      # T.id = ANY(...)
.WHERE(arr.length(T.items) > 0)

# Full-text — to_tsvector / web_query / matches / rank
.WHERE(fts.matches(
    fts.to_tsvector(T.body),
    fts.web_query(user_input)
))
.ORDER_BY(
    fts.rank(fts.to_tsvector(T.body), fts.web_query(user_input)),
    DESC=True,
)

For dict-to-JSONB autoadaptation in psycopg, register the dumper once at startup:

import psycopg
from psycopg.types.json import JsonbDumper
psycopg.adapters.register_dumper(dict, JsonbDumper)

DISTINCT and DISTINCT ON

# Plain DISTINCT — deduplicate every selected row.
await cygnet.SELECT(db, T.country).FROM(T).DISTINCT()

# DISTINCT ON (cols) — PG-specific: keep one row per distinct value of
# the listed columns, picked according to ORDER BY.
await (
    cygnet.SELECT(db, T.country, T.name)
    .DISTINCT_ON(T.country)
    .FROM(T)
    .ORDER_BY(T.country, T.name)   # determines which row wins per country
)

Set operations

# UNION dedupes; UNION_ALL preserves duplicates.
combined = await (
    cygnet.SELECT(db, A.name).FROM(A)
    .UNION(cygnet.SELECT(db, B.name).FROM(B))
    .ORDER_BY(cygnet.lit("name"))    # applies to the COMPOUND result
    .LIMIT(100)
)

# Other set ops: INTERSECT / INTERSECT_ALL / EXCEPT_ / EXCEPT_ALL.
# (EXCEPT_ has a trailing underscore because `except` is a Python keyword.)
diff = await (
    cygnet.SELECT(db, A.name).FROM(A)
    .EXCEPT_(cygnet.SELECT(db, B.name).FROM(B))
)

Streaming large result sets

# `async for ...` instead of `await ...`: rows arrive one at a time
# from a server-side portal cursor, so the full result set never lives
# in process memory.
async with cygnet.transaction(db) as tx:
    async for entry in (
        cygnet.SELECT(tx)
        .FROM(LogTable)
        .WHERE(LogTable.account_id == 42)
        .ORDER_BY(LogTable.id)
        .stream()
    ):
        process(entry)

PostgreSQL portal cursors require a transaction (or autocommit off), so streaming is typically wrapped in cygnet.transaction(db). The db adapter must implement an async stream(sql, params) method; psycopg's cursor.stream() is the reference implementation — see cygnet/psycopg_db.py.

If you break out of the async for early, wrap the loop in contextlib.aclosing(...) (or rely on the enclosing cygnet.transaction(db) block, whose commit/rollback drops the server-side portal) so the cursor closes deterministically — a bare break leaves cleanup to garbage collection, which can't run an async close.

Window functions

import cygnet.functions as f

# `row_number() OVER (PARTITION BY dept ORDER BY salary DESC)`
rn = f.row_number().OVER(
    partition_by=[EmployeeTable.dept],
    order_by=[(EmployeeTable.salary, "DESC")],
)
rows = await (
    cygnet.SELECT(db, EmployeeTable.name, EmployeeTable.dept, rn)
    .FROM(EmployeeTable)
)

# Aggregates work as windows too
avg_salary = f.avg(EmployeeTable.salary).OVER(
    partition_by=[EmployeeTable.dept]
)

# LAG / LEAD pull from neighbouring rows
prev = f.lag(EmployeeTable.salary, 1).OVER(order_by=[EmployeeTable.id])

# Frames are passed as raw SQL strings (interpolated verbatim — trusted)
running_sum = f.sum(T.amount).OVER(
    order_by=[T.id],
    frame="ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW",
)

The available curated window functions: row_number, rank, dense_rank, percent_rank, cume_dist, ntile, lag, lead, first_value, last_value, nth_value. Anything else reaches via cygnet.fn("name")(...).OVER(...).

CTEs (WITH clauses)

# Build the inner SELECT, wrap it in a CTE, then reference it like a table.
active = cygnet.cte(
    "active",
    cygnet.SELECT(db, AccountTable.id, AccountTable.name)
        .FROM(AccountTable)
        .WHERE(AccountTable.status == "active"),
)
results = await (
    cygnet.SELECT(db, active.name, LogTable.message)
    .WITH(active)
    .FROM(active)
    .LEFT_JOIN(LogTable, ON=active.id == LogTable.account_id)
)

CTE column names are inferred from the inner SELECT's ColumnProxy arguments. For inner SELECTs that use opaque expressions (cygnet.fn, cygnet.lit, …), pass columns=[…] explicitly:

counts = cygnet.cte(
    "counts",
    cygnet.SELECT(db, AccountTable.id, cygnet.fn("count")(LogTable.id))
        .FROM(AccountTable)
        .LEFT_JOIN(LogTable, ON=AccountTable.id == LogTable.account_id)
        .GROUP_BY(AccountTable.id),
    columns=["id", "n"],
)

Multiple CTEs compose, either via repeated .WITH() or in one call:

.WITH(active, recent)
.WITH(active).WITH(recent)

Recursive CTEs require explicit columns (so the recursive step can reference them) and an anchor + step assigned after construction:

counter = cygnet.recursive_cte("counter", columns=["n"])
counter.anchor = cygnet.SELECT(db, cygnet.lit("1"))
counter.step = (
    cygnet.SELECT(db, counter.n + 1)
    .FROM(counter)
    .WHERE(counter.n < 10)
)

rows = await cygnet.SELECT(db, counter.n).WITH(counter).FROM(counter)
# [(1,), (2,), …, (10,)]

When a recursive CTE appears in a WITH(…) list, the rendered SQL uses WITH RECURSIVE for the whole list (PG's syntax requirement).

Transactions

async with cygnet.transaction(db) as tx:
    await cygnet.INSERT(tx).INTO(AccountTable).VALUES(acc)
    await cygnet.INSERT(tx).INTO(LogTable).VALUES(entry)
# commits on clean exit, rolls back on exception

Nested transaction blocks transparently use SAVEPOINT:

async with cygnet.transaction(db) as tx:
    await cygnet.save(tx, acc)
    async with cygnet.transaction(tx) as tx2:   # SAVEPOINT
        await cygnet.save(tx2, risky_entry)      # RELEASE or ROLLBACK TO

Concurrency caveat. cygnet.transaction toggles a _in_transaction flag on the db handle to detect nesting, and that flag is not task-local. A single db connection must not be shared across concurrent asyncio tasks. psycopg connections are themselves not task-safe, so the recommended pattern is one connection per task — typically by acquiring from a pool inside each task. Fresh connections must start with _in_transaction = False.

Cygnet actively detects cross-task misuse: the outermost __aenter__ records the owning asyncio.current_task() on the db, and a nested __aenter__ from a different task raises RuntimeError rather than silently SAVEPOINTing inside the other task's transaction. The guard is best-effort — it only fires when the outer layer uses cygnet.transaction rather than an externally-managed BEGIN.

Annotations

Annotation	Meaning
cygnet.DBKey	Primary key assigned by the database (SERIAL / IDENTITY). Omitted on INSERT when None; populated via RETURNING. Incompatible with frozen=True.
cygnet.AppKey	Primary key assigned by the application (e.g. UUID). Must never be None.
cygnet.Column("col_name")	Override the column name for a field.
cygnet.ForeignKey(Target)	Mark a field as a foreign key referencing Target's primary key. Enables FOLLOW / LEFT_FOLLOW join sugar and cygnet.follow(). Composite PKs are not supported.
@cygnet.table("table_name")	Override the table name for a dataclass (default: classname.lower() + "s").

The db object

CYGNET does not manage connections. Pass any object that conforms to cygnet.DBAdapter — a @runtime_checkable Protocol declared in cygnet/expression.py and re-exported at the package root. Required members:

class DBAdapter(Protocol):
    _in_transaction: bool        # False on fresh adapter; toggled by cygnet.transaction
    _transaction_task: Any       # Cygnet-managed task-locality stash; init to None

    async def execute(self, sql: str, params: list | None = None) -> list[tuple]: ...
    async def execute_one(self, sql: str, params: list | None = None) -> tuple | None: ...

Optional methods (duck-typed via hasattr, not on the Protocol):

# Only required for SelectBuilder.stream():
async def stream(self, sql: str, params: list | None = None) -> AsyncIterator[tuple]: ...

# Only required for DEFAULT-aware INSERT codegen (None-valued columns
# with a schema DEFAULT omitted from INSERT, refreshed via RETURNING):
async def column_defaults(self, table_name: str) -> set[str]: ...

Because DBAdapter is runtime_checkable, custom adapters can verify conformance with a plain isinstance(my_db, cygnet.DBAdapter) check before shipping. The reference PsycopgDB adapter implements both required AND both optional methods.

Reference psycopg3 adapter

Install with the [psycopg] extra (see Installation), then:

import psycopg
from cygnet.psycopg_db import PsycopgDB

conn = await psycopg.AsyncConnection.connect("postgresql://...")
db = PsycopgDB(conn)
accounts = await cygnet.SELECT(db).FROM(AccountTable)

PsycopgDB translates Cygnet's $N placeholders to psycopg's %s, implements all four protocol methods (including stream() via portal cursor), and tracks _in_transaction for cygnet.transaction().

Importing cygnet.psycopg_db without the extra installed raises a clear ImportError pointing at the install command — Cygnet's core itself stays driver-free, so a project shipping a custom adapter never pulls psycopg.

Reference asyncpg adapter

Install with the [asyncpg] extra (see Installation), then:

import asyncpg
from cygnet.asyncpg_db import AsyncpgDB

conn = await asyncpg.connect("postgresql://...")
db = AsyncpgDB(conn)
accounts = await cygnet.SELECT(db).FROM(AccountTable)

AsyncpgDB satisfies the same DBAdapter protocol as PsycopgDB. asyncpg speaks libpq's native $N placeholders — exactly what Cygnet emits — so there's no $N→%s translation. This is a first cut: it implements the two required methods but neither optional one (stream() nor column_defaults()), so a None-valued non-PK column with a schema DEFAULT is inserted as an explicit NULL rather than letting the DEFAULT fire — a SERIAL/DBKey primary key is unaffected. The rationale is in THEORY.md.

Connection pooling

from psycopg_pool import AsyncConnectionPool
from cygnet.psycopg_db import PsycopgDB

pool = AsyncConnectionPool("postgresql://...")
await pool.open()

async def handler(...):
    # One connection per task — never share a PsycopgDB across
    # concurrent tasks, since `_in_transaction` is per-instance and
    # not task-local.
    async with pool.connection() as conn:
        db = PsycopgDB(conn)
        async with cygnet.transaction(db):
            await cygnet.save(db, obj)

Strict typing & IDE autocomplete

Out of the box, cygnet.Table(Account) returns TableProxy[Account], so cygnet.get(db, AccountTable, id=1) correctly types as Account | None. What mypy can't infer without help is the per-field shape: AccountTable.name resolves to ColumnProxy (effectively Any), so a typo like AccountTable.nmae doesn't error at type-check time and IDEs don't autocomplete attribute names.

For projects that want strict typing, cygnet.stubs generates a hand-pasteable TYPE_CHECKING block:

$ python -m cygnet.stubs myapp.models

Output (paste into myapp/models.py, replacing the original XTable = cygnet.Table(X) lines):

from typing import TYPE_CHECKING

if TYPE_CHECKING:
    from cygnet.proxy import ColumnProxy, TableProxy

    class _AccountTable(TableProxy[Account]):
        id: ColumnProxy[int]
        name: ColumnProxy[str]
        email: ColumnProxy[str]

    AccountTable: _AccountTable
else:
    import cygnet
    AccountTable = cygnet.Table(Account)

Mypy now sees AccountTable.name as ColumnProxy[str]; runtime is unchanged. Regenerate after schema changes.

PG type adapters

psycopg3 handles most PG types natively (uuid.UUID, decimal.Decimal, datetime.datetime/date/time, arrays of those, etc.). A few types need one-line registration at app startup:

import psycopg
from psycopg.types.json import JsonbDumper

# Plain Python dicts -> JSONB (Cygnet.jsonb helpers expect this).
psycopg.adapters.register_dumper(dict, JsonbDumper)

For ranges (int4range, tstzrange, …), psycopg provides Range and DateTimeTzRange etc. as native Python types; no extra registration needed. See psycopg3's docs for the full list.

Design principles

No magic — no base class, no metaclass, no implicit queries; you call it, it runs. SQL keywords are uppercase methods, Python utilities are lowercase, PostgreSQL-specific features are used directly, and you bring your own connection and transaction lifecycle. The reasoning behind each of these — and the alternatives that were rejected — is in THEORY.md.

Development

Development requires uv as the package manager — install it via your platform's package manager (Homebrew / apt / pipx install uv) before the steps below. just bootstrap runs uv sync --extra dev which honours the checked-in uv.lock for reproducible installs.

just bootstrap       # create .venv via uv sync, install dev dependencies (locked)
just check           # fmt + lint + typecheck + unit tests
just test-all        # full suite including integration (requires Docker)
just pg-psql         # open psql against the test container

Benchmarks

The bench/ suite tracks Cygnet's performance over time (advisory in CI — never blocks merge) and gives a feel for ORM overhead vs total wall time.

just bootstrap-bench    # one-time: install pytest-benchmark + Django + SQLAlchemy
just bench              # render + overhead benchmarks (no DB, ~5s)
just bench-e2e          # end-to-end against a fresh Docker PG (~30s)
just bench-all          # everything, JSON output for CI artifact

Three layers of measurement, each focused on a different cost:

• bench/test_render.py — pure SQL generation against FakeDB. Sub-microsecond per call; isolates the cost of building the AST and rendering it to a string.
• bench/test_overhead.py — full Cygnet path through FakeDB, including row-to-object hydration. Catches regressions in the executor and mapper.
• bench/test_e2e.py — real PG via PsycopgDB. Total wall time including round-trip.

CI runs all three on every push and PR, uploads the JSON as an artifact, and posts a summary table in the job output. Regressions are informational; the job is continue-on-error: true so a noisy benchmark never wedges merging.

On pull_request events, the bench job additionally downloads main's last successful bench-results artifact and renders a per-benchmark delta table in the same step summary. Slowdowns greater than 15% appear bold; runner-noise-sized changes (±10%) stay plain text. The comparison gracefully no-ops when no baseline exists yet (first PR after a fresh repo, expired retention, etc.).

Cross-ORM comparison

bench/comparison/test_comparison.py runs the same operations against the same PG schema through five columns: Cygnet and SQLAlchemy 2 (async session) each on both psycopg and asyncpg, plus Django (sync ORM). Five operation classes × five columns = twenty-five side-by-side benchmarks — isolating two axes at once, ORM overhead (Cygnet vs SA vs Django) and driver overhead (psycopg vs asyncpg) within an ORM. The operations: SELECT-by-PK, SELECT-all-100, a 1000-row JOIN (Cygnet FOLLOW / SQLAlchemy join / Django select_related), INSERT one, and bulk INSERT 50.

Each ORM is benchmarked in its idiomatic mode — Cygnet and SA in async, Django in sync — so the numbers reflect what real applications see. SA's connection pool is clamped to a single connection (pool_size=1, max_overflow=0) to match Cygnet's single connection and Django's per-request connection, so the deltas measure ORM overhead rather than connection management.

Informal results

A single internally-consistent run on an Apple M3 Max (Python 3.13, a local Dockerised PostgreSQL 16 started with fsync=off synchronous_commit=off full_page_writes=off, single connection, no pooling). These are informal, laptop-grade, single-run numbers — read them as orders of magnitude, not a benchmark league table. Median µs per operation; bold marks the fastest cell in each row.

Operation	Cygnet/psycopg	Cygnet/asyncpg	SA/psycopg	SA/asyncpg	Django
SELECT by primary key (1 row)	757	566	859	1,023	963
SELECT all (100 rows)	569	559	893	869	644
JOIN posts→accounts (1000 rows)	3,456	3,511	6,069	5,919	6,015
INSERT one row	535	501	2,318	1,080	220
Bulk INSERT (50 rows)	469	394	1,934	1,633	705

The SELECT-by-PK row is left unbolded on purpose — it is within noise (see the first caveat below), so picking a "winner" there would over-read the spread.

The honest reading is "Cygnet's overhead is small," not "Cygnet is fastest." A thin SQL builder should sit close to the driver — there's little between your call and the wire. Cygnet (on either driver) is fastest-or-tied on the reads and the JOIN; on writes the only ORM that reads faster is Django on INSERT-one, and that bold cell is an artifact — see the async-tax caveat below. The caveats matter as much as the numbers:

• The single-row read row is noisy. SELECT-by-PK has wide spread (IQR ~560 µs on the Cygnet/psycopg cell, ~416 µs on SA/psycopg). The lean cells in that row — both Cygnet drivers and Django — are within run-to-run noise of one another; don't read fine-grained ordering into them, and in particular don't read Cygnet/psycopg as meaningfully slower than Cygnet/asyncpg here.
• The database is deliberately fast. fsync=off, synchronous_commit=off, and full_page_writes=off on a localhost Docker PG make commits nearly free, which maximises the visible share of ORM overhead. Against a real, durable, networked database these per-call gaps shrink to a sliver of total query latency.
• One connection, no pooling — these are ORM-overhead numbers, not a connection-management comparison.
• The async tax favors Django (a harness artifact). Cygnet and SA each pay one event-loop entry (loop.run_until_complete) per measured op; Django's sync path pays none. This inflates the async ORMs' absolute numbers on cheap ops — it's why Django reads as fastest on INSERT-one (220 µs) despite issuing the same INSERT … RETURNING. Django is not faster at the actual insert; strip the loop spin and Cygnet is at or below it. A real async app amortises one loop across many awaits.
• SA's read columns understate Cygnet's lead. The SA session is reused across rounds and the read ops never commit/expire, so Session.get(pk) serves from the identity map without SQL after round 1, and select() execution skips re-hydrating already-loaded rows. SA is doing strictly less work on the read rows. The identity map is a real SQLAlchemy feature; Cygnet has no identity map by design.
• SA's writes aren't strictly like-for-like. Those cases pay AsyncSession construction + a unit-of-work flush + BEGIN/COMMIT vs Cygnet's single autocommit statement; the large SA insert ratios are session/UoW idiom, not raw statement cost.
• The driver axis (asyncpg vs psycopg) crosses over by result-set shape. Within Cygnet, asyncpg wins on writes and single-row reads, is roughly flat at 100 rows, and is a touch slower on the 1000-row JOIN. That crossover is the asyncpg adapter's Approach A: it converts each asyncpg Record to a tuple at the boundary (keeping hydration on Cygnet's fast positional path), a fixed per-row cost that only pays off where there's little per-row hydration. See THEORY.md for the rationale. The SA driver swap is mixed/neutral — its ORM machinery dominates the driver.
• Object richness is a real trade-off, not a measurement error. Cygnet returns inert dataclasses; SA and Django return change-tracked instances. Cheaper objects are part of why Cygnet is faster.
• Your hardware will disagree. Re-run it and see.

Skipped automatically when CYGNET_TEST_DSN is unset; reproduce with:

just bootstrap-bench      # one-time: Django + SQLAlchemy + pytest-benchmark
just bench-all            # spins up a throwaway Docker PG, runs everything
# …or point it at your own database:
CYGNET_TEST_DSN=postgresql://… pytest bench/comparison/ --benchmark-only

License

MIT