File size: 9,962 Bytes
942050b 4e1037f 942050b 4e1037f 942050b 4e1037f 942050b 4e1037f 942050b 4e1037f 942050b 4e1037f 942050b 4e1037f 942050b 4e1037f 942050b 4e1037f 942050b 4e1037f 942050b 4e1037f 942050b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 | """Read-only connection helpers for target databases.
The NLβSQL pipeline never owns write privileges on a target DB. Defences:
- Postgres: dedicated role with `default_transaction_read_only=on`, see
`scripts/sql/postgres_init.sql`.
- SQLite: `mode=ro` URI passed via a SQLAlchemy creator (URL form does not
carry through cross-platform; creator gives us full control over path
encoding) plus `PRAGMA query_only=ON` as a belt-and-braces guard.
"""
from __future__ import annotations
import sqlite3
import threading
import time
from collections.abc import Iterator
from contextlib import contextmanager
from dataclasses import dataclass
from pathlib import Path
from typing import Any, Literal
from sqlalchemy import Connection, Engine, create_engine, text
from sqlalchemy.engine import make_url
from sqlalchemy.exc import DBAPIError
Dialect = Literal["sqlite", "postgresql"]
# One Engine (and therefore one connection pool) per spec, process-wide.
# `make_engine` is called on the hot path β once per /ask, once per eval
# question β and a fresh `create_engine` there builds a fresh pool that is never
# disposed: each question opened a new Postgres backend and left the old pool to
# the garbage collector. Under sustained load that walks straight into
# max_connections. Engines are thread-safe and designed to be long-lived; it is
# Connections that are not.
_ENGINE_CACHE: dict[DatabaseSpec, Engine] = {}
_ENGINE_LOCK = threading.Lock()
@dataclass(frozen=True, slots=True)
class DatabaseSpec:
"""Connection target.
For SQLite, `url` is the absolute filesystem path to the .sqlite file.
For Postgres, `url` is a standard libpq DSN (`postgresql://...`).
"""
id: str
dialect: Dialect
url: str
description: str = ""
def make_engine(self) -> Engine:
"""Return this spec's engine, building it once and pooling it thereafter."""
return _engine_for(self)
@dataclass(frozen=True, slots=True)
class QueryResult:
rows: list[tuple[Any, ...]]
columns: list[str]
row_count: int
truncated: bool
elapsed_ms: float
def _build_engine(spec: DatabaseSpec) -> Engine:
if spec.dialect == "sqlite":
return _build_sqlite_readonly_engine(Path(spec.url))
if spec.dialect == "postgresql":
return _build_postgres_readonly_engine(spec.url)
raise ValueError(f"unsupported dialect: {spec.dialect}")
def _normalise_pg_driver(url: str) -> str:
"""Force the psycopg (v3) driver for bare ``postgresql://`` DSNs.
SQLAlchemy resolves an unqualified ``postgresql://`` to the psycopg2 driver,
which is not installed here (only psycopg 3 is). Left unqualified, every
Postgres connection would die with ModuleNotFoundError β which is exactly
why the Postgres path had never actually run. Explicit ``+psycopg`` /
``+psycopg2`` DSNs are respected as-is.
"""
parsed = make_url(url)
if parsed.drivername == "postgresql":
parsed = parsed.set(drivername="postgresql+psycopg")
return parsed.render_as_string(hide_password=False)
def _build_postgres_readonly_engine(url: str) -> Engine:
"""Engine whose every transaction is genuinely READ ONLY.
The read-only guarantee comes from SQLAlchemy's ``postgresql_readonly``
execution option, which sets the connection read-only at transaction start
(the correct point). An earlier implementation issued
``SET default_transaction_read_only = on`` from inside an already-open
transaction β Postgres fixes a transaction's read-only status at BEGIN, so
that statement was a no-op for the current transaction and the "layer 1"
read-only defence never actually engaged.
"""
return create_engine(
_normalise_pg_driver(url),
future=True,
pool_pre_ping=True,
execution_options={"postgresql_readonly": True},
)
def _build_sqlite_readonly_engine(path: Path) -> Engine:
if not path.is_absolute():
path = path.resolve()
file_uri = path.as_uri() + "?mode=ro"
def _creator() -> sqlite3.Connection:
conn = sqlite3.connect(file_uri, uri=True, check_same_thread=False)
conn.execute("PRAGMA query_only = ON")
return conn
return create_engine("sqlite://", creator=_creator, future=True)
def _engine_for(spec: DatabaseSpec) -> Engine:
"""Cached engine for a spec. DatabaseSpec is frozen, so it keys the cache."""
engine = _ENGINE_CACHE.get(spec)
if engine is not None:
return engine
with _ENGINE_LOCK:
engine = _ENGINE_CACHE.get(spec)
if engine is None:
engine = _build_engine(spec)
_ENGINE_CACHE[spec] = engine
return engine
def dispose_engines() -> None:
"""Drop every pooled engine. For tests, and for callers that swap a DB file
underneath a spec whose url (the cache key) has not changed."""
with _ENGINE_LOCK:
for engine in _ENGINE_CACHE.values():
engine.dispose()
_ENGINE_CACHE.clear()
def connect(spec: DatabaseSpec) -> Engine:
"""Build (or reuse via SQLAlchemy pool) an engine for a DB spec."""
return spec.make_engine()
@contextmanager
def execute_readonly(
engine: Engine,
sql: str,
*,
statement_timeout_ms: int = 30_000,
row_cap: int = 10_000,
) -> Iterator[QueryResult]:
"""Run a SELECT-only query with hard timeout and row cap.
Caller must have already validated `sql` through the AST guard. This
function enforces operational limits, not correctness or safety.
"""
started = time.perf_counter()
with engine.connect() as conn, _runtime_limits(conn, statement_timeout_ms):
# Run on a DBAPI cursor with NO parameters, because every parameterised
# path mangles a literal that real SQL contains:
# - SQLAlchemy `text()` reads `:__` as a bind parameter, breaking gold
# like `LIKE '_:%:__.___'` (BIRD qids 959/989/990, formula_1 times).
# - `exec_driver_sql` passes an (empty) parameter set to the driver, so
# psycopg's pyformat parser scans for placeholders and rejects `%`:
# `LIKE '%variance%'` β "only '%s', '%b', '%t' are allowed as
# placeholders". That killed LIKE filters *and* the `%` modulo
# operator on Postgres β see tests/test_postgres_readonly.py.
# A cursor executed with no parameters interprets neither, on either driver.
cursor = conn.connection.cursor()
try:
cursor.execute(sql)
columns = [str(d[0]) for d in cursor.description or ()]
rows = cursor.fetchmany(row_cap + 1)
except Exception as raw_exc:
# Going through the raw cursor also means driver-native exceptions
# (sqlite3.OperationalError, psycopg.Error) instead of SQLAlchemy's.
# Callers classify failures on sqlalchemy.exc types β timeouts on
# OperationalError, everything else on SQLAlchemyError (see
# execution/runner.py) β so re-wrap and keep that contract intact.
raise DBAPIError.instance(
sql,
None,
raw_exc,
engine.dialect.loaded_dbapi.Error,
dialect=engine.dialect,
) from raw_exc
finally:
cursor.close() # release server-side resources / drain remaining rows
truncated = len(rows) > row_cap
if truncated:
rows = rows[:row_cap]
elapsed_ms = (time.perf_counter() - started) * 1000.0
yield QueryResult(
rows=[tuple(r) for r in rows],
columns=columns,
row_count=len(rows),
truncated=truncated,
elapsed_ms=elapsed_ms,
)
@contextmanager
def _runtime_limits(conn: Connection, statement_timeout_ms: int) -> Iterator[None]:
"""Arm the per-query timeout, and disarm it again on the way out."""
dialect = conn.engine.dialect.name
if dialect == "postgresql":
# statement_timeout takes effect immediately for subsequent statements in
# this session, so setting it here (inside the transaction) is correct.
# Read-only is NOT set here β it is applied at transaction start by the
# engine's postgresql_readonly execution option (see
# _build_postgres_readonly_engine); issuing it here would be a no-op.
conn.execute(text(f"SET statement_timeout = {int(statement_timeout_ms)}"))
yield
return
raw = conn.connection.driver_connection if dialect == "sqlite" else None
if not isinstance(raw, sqlite3.Connection):
yield
return
# SQLite has no per-query timeout knob in the URL API; use the connection
# progress handler at the dbapi level.
_install_sqlite_timeout(raw, statement_timeout_ms / 1000.0)
try:
yield
finally:
# Engines are pooled now, so this connection goes back to the pool and on
# to the next caller β including schema introspection, which never arms a
# handler of its own. A handler left behind carries an already-expired
# deadline and would interrupt that caller's first non-trivial query.
raw.set_progress_handler(None, 0)
def _install_sqlite_timeout(conn: sqlite3.Connection, seconds: float) -> None:
deadline = time.monotonic() + seconds
def _interrupt() -> int:
return 1 if time.monotonic() > deadline else 0
# progress handler is invoked every N VM ops; 1000 keeps overhead trivial.
conn.set_progress_handler(_interrupt, 1000)
def sqlite_url_readonly(path: Path) -> str:
"""Return the absolute path used as DatabaseSpec.url for SQLite specs.
We store the bare path (not a full SQLAlchemy URL) because read-only mode
is applied via a creator function in `_build_sqlite_readonly_engine` β
SQLAlchemy's URL builder does not carry the SQLite `mode=ro` URI cleanly
across Windows and POSIX path encodings.
"""
return str(path.resolve())
|