Distributed Locks
Technical reference for eventsource.ports.locks and eventsource.adapters.postgresql.locks: the PostgreSQLLockManager advisory-lock manager, the LockInfo value object, the migration_lock_key helper, and the two lock-specific exceptions.
eventsource.locks -- the pre-slice-A import path for all of the above -- no longer exists (ADR 0030). Importing it raises ModuleNotFoundError, with no deprecation shim. Update imports to the paths this page documents.
The package provides mutual exclusion across processes — one holder at a time for a given string key, regardless of which application instance asks. It is built entirely on PostgreSQL session-level advisory locks (pg_advisory_lock, pg_try_advisory_lock, pg_advisory_unlock), which are application-level locks independent of table and row locks. A lock persists for the lifetime of the database session that holds it and is released automatically when that session ends, including on an unexpected disconnect.
The names split across three modules along ring boundaries (ADR 0029):
| Name | Kind | Lives in |
|---|---|---|
PostgreSQLLockManager |
Lock manager class | eventsource.adapters.postgresql.locks |
LockInfo |
Frozen dataclass describing an acquired lock | eventsource.ports.locks |
migration_lock_key |
Lock-key naming helper | eventsource.ports.locks |
LockAcquisitionError |
Raised when a lock cannot be acquired | eventsource.ports.exceptions |
LockNotHeldError |
Raised when releasing a lock this manager does not hold | eventsource.ports.exceptions |
The package's primary in-tree consumer is the live migration tooling (eventsource.application.migration), where a lock guards cutover so that only one instance can flip a tenant's event store at a time — hence the migration_lock_key convenience helper. Nothing about the manager is migration-specific, though: any operation that must not run concurrently across instances can use it.
Overview
PostgreSQLLockManager is the only lock manager in the package. It is constructed from a SQLAlchemy async_sessionmaker[AsyncSession] and exposes two acquisition styles plus a small set of inspection and cleanup operations:
| Member | Signature | Purpose |
|---|---|---|
acquire |
acquire(key, *, timeout=None, retry_interval=0.1) |
Async context manager; yields LockInfo, always releases on exit |
try_acquire |
await try_acquire(key) |
One-shot, non-blocking; returns LockInfo or None |
release |
await release(key) |
Releases a lock taken with try_acquire; raises LockNotHeldError if not held |
is_held |
await is_held(key) |
Whether this manager currently holds key |
release_all |
await release_all() |
Releases every lock this manager holds; returns the count |
held_lock_count |
property | Number of locks currently held (plain int, not async) |
Locks are named by arbitrary strings. The manager hashes each key with SHA-256 and truncates the digest to 63 bits to produce the bigint lock ID that PostgreSQL's advisory-lock functions actually take, so callers never deal with numeric IDs directly — LockInfo.lock_id exposes the derived value only for diagnostics.
The two acquisition styles differ in who is responsible for cleanup. acquire() is the preferred form: it is an @asynccontextmanager, and its finally block releases the lock whether the body returns normally or raises. try_acquire() returns immediately — LockInfo when it won the lock, None when another session holds it — and leaves release entirely to the caller, who must pair it with release(key) in a try/finally.
Waiting behaviour is controlled by timeout on acquire(). With the default timeout=None the manager issues a single blocking pg_advisory_lock and waits indefinitely. With a timeout set it polls pg_try_advisory_lock every retry_interval seconds until the deadline, then raises LockAcquisitionError. try_acquire() never waits and never raises for contention — it simply returns None.
Because advisory locks are session-scoped in PostgreSQL, the manager checks out a dedicated AsyncSession per held lock and keeps it open for as long as the lock is held, closing it on release. That is the single most important operational consequence of this design: N concurrently held locks occupy N connections from the pool. It also means a crashed process releases its locks automatically, since the server drops the advisory lock when the backing session disconnects.
The manager tracks held locks in an in-process dictionary guarded by an asyncio.Lock, so is_held, release, and release_all reflect only what this manager instance holds — not locks held by other processes or by another manager against the same database. Every acquire and release is wrapped in an OpenTelemetry span when tracing is enabled, and both emit DEBUG log records.
PostgreSQL-Only Constraint
eventsource.ports.locks now defines DistributedLock and LockRegistry -- small Protocols describing the shape of the dependency (acquire/release/ is_held; bulk release and count), split along the two real consumer groups (ADR 0029). That is narrower than a backend-agnostic distributed-locking guarantee: the Protocols say nothing about cross-process exclusion, release on crash, or fairness, because only one implementation offers those properties. PostgreSQLLockManager is the only production implementation and requires a PostgreSQL connection; InMemoryLockManager (eventsource.adapters.memory.locks) conforms to the same Protocols for single-process testing only -- see InMemoryLockManager below. There is no SQLite variant.
The constraint is structural, not a matter of configuration. PostgreSQLLockManager issues three PostgreSQL-specific statements directly as raw SQL through SQLAlchemy's text():
| Operation | Statement |
|---|---|
Blocking acquire (timeout=None) |
SELECT pg_advisory_lock(:lock_id) |
Non-blocking acquire (try_acquire, and each poll when timeout is set) |
SELECT pg_try_advisory_lock(:lock_id) |
| Release | SELECT pg_advisory_unlock(:lock_id) |
These functions exist only in PostgreSQL. Nothing in the manager branches on dialect, and no fallback path exists. Pointing the session_factory at a SQLite or other non-PostgreSQL engine does not degrade gracefully — the first acquire() or try_acquire() call fails when the database rejects the unknown function. In acquire() and its timeout loop that failure is caught and re-raised as LockAcquisitionError with a reason of "Database error: ..."; try_acquire() closes the session, logs at ERROR, and re-raises the underlying driver exception unchanged.
Two further consequences of the PostgreSQL binding are worth stating explicitly:
- The 63-bit lock ID exists because of PostgreSQL.
pg_advisory_lockand friends take a signedbigint, so the manager masks its SHA-256 digest to 63 bits (& 0x7FFFFFFFFFFFFFFF). See Key-to-Lock-ID Derivation. - Lock lifetime is PostgreSQL session lifetime. The manager holds one
AsyncSessionopen per lock precisely because advisory locks are session-scoped on the server. See Session and Connection Pool Model.
No optional-dependency guard
The package follows the library's optional-backend convention only in spirit, not in mechanics. There is no try/except ImportError wrapper and no POSTGRES_AVAILABLE-style flag to check at runtime, because the module imports nothing beyond the standard library, sqlalchemy, and eventsource.observability — and sqlalchemy is a core dependency. asyncpg is not imported here at all; the driver is whatever the caller's async_sessionmaker was built with. import eventsource.adapters.postgresql.locks therefore always succeeds on a base install, and PostgreSQL availability is discoverable only by attempting an acquisition.
Note also that the lock names are not re-exported from the top-level eventsource package. They are not reachable via eventsource. Import them from eventsource.ports.locks / eventsource.adapters.postgresql.locks / eventsource.domain.exceptions (see Import Surface).
Effect on migration cutover
Because MigrationCoordinator.cutover delegates coordination to a PostgreSQLLockManager, cutover inherits this constraint: a PostgreSQL database must be reachable to serialize it across instances, independent of which stores are being migrated between. The coordinator's lock_manager argument is optional at construction (lock_manager: PostgreSQLLockManager | None = None), so a coordinator can be built for sync-only work without one, but calling cutover without a lock manager raises MigrationError rather than proceeding unguarded.
If you need mutual exclusion in a deployment with no PostgreSQL instance, this package will not provide it — use an external coordination service, or, for single-process cases where cross-instance exclusion is not actually required, an asyncio.Lock.
`InMemoryLockManager` (test-only)
Lead with what it does not guarantee. InMemoryLockManager (eventsource.adapters.memory.locks) excludes only coroutines running in one asyncio event loop in one process:
- No cross-process exclusion. It is a dict guarded by an
asyncio.Condition, not a database primitive. Two OS processes never contend against the sameInMemoryLockManagerinstance. - No release on crash. There is no server session to drop a lock when the holder disappears. A process that dies mid-critical-section leaves the lock held until the manager itself is garbage collected.
- No fairness. Waiters are woken in whatever order
asyncio.Conditionhappens to wake them; there is no FIFO queue. - Single event loop, single process. Construct one per test, per loop. Sharing an instance across loops is unsupported, same as
PostgreSQLLockManager.
It conforms to the same DistributedLock / LockRegistry Protocols as PostgreSQLLockManager -- same acquire/try_acquire/release/is_held signatures, same LockAcquisitionError / LockNotHeldError on failure, and the same key-to-lock-id hashing (LockInfo.lock_id matches what PostgreSQLLockManager would derive for the same key) -- so code written against the port can be tested against it without a database:
from eventsource.adapters.memory.locks import InMemoryLockManager
lock_manager = InMemoryLockManager(holder_id="test-worker")
async with lock_manager.acquire("cutover:tenant-123") as lock_info:
... # exercise code that needs a DistributedLock
Use it in unit tests that need a DistributedLock and nothing more. Use PostgreSQLLockManager anywhere two processes must actually coordinate.
Import Surface
The five public names split across three modules by ring (ADR 0029):
from eventsource.ports.locks import LockInfo, migration_lock_key
from eventsource.ports.exceptions import LockAcquisitionError, LockNotHeldError
from eventsource.adapters.postgresql.locks import PostgreSQLLockManager
eventsource.locks (the pre-slice-A path) no longer exists (ADR 0030); importing it raises ModuleNotFoundError.
Not exported from the top-level package
None of these names appear in the top-level eventsource namespace. from eventsource import PostgreSQLLockManager raises ImportError. This is a deliberate departure from the library's usual convention of re-exporting the public API from eventsource/__init__.py, and it is consistent with the locks' PostgreSQL-only production nature — they are a backend-specific tool rather than part of the core surface.
The exceptions live in `eventsource.ports.exceptions` now
LockAcquisitionError and LockNotHeldError subclass EventSourceError (ADR 0029, widening only) and live in eventsource.ports.exceptions (ADR 0041 moved them there from eventsource.domain.exceptions, alongside twelve other infrastructure exceptions, because a lock-acquisition failure is a port-contract concern, not a domain one). Two consequences for calling code:
- Import them from
eventsource.ports.exceptions, not from the locks modules and not fromeventsource.domain.exceptions(that import path now raisesImportErrorfor these two names -- no shim). except EventSourceErrornow catches a lock failure too. Every existingexcept LockAcquisitionErrorandexcept Exceptionstill catches exactly as before; the newly-catching clause isexcept EventSourceError, which caught nothing lock-related prior to this change.
In-tree, eventsource.application.migration.cutover follows this pattern, importing LockAcquisitionError from eventsource.ports.exceptions and migration_lock_key from eventsource.ports.locks.
Typing-only imports
PostgreSQLLockManager is referenced under if TYPE_CHECKING: in eventsource.application.migration.cutover and eventsource.application.migration.coordinator, which is the right approach for annotating a lock_manager parameter without pulling eventsource.adapters.postgresql.locks into module import at runtime. from __future__ import annotations is already in effect in those modules, so the deferred import costs nothing at runtime.
What the import costs
eventsource.ports.locks is pure: stdlib only (contextlib, dataclasses, datetime, typing, uuid), no sqlalchemy. Importing eventsource.adapters.postgresql.locks pulls in the standard library (asyncio, hashlib, logging, contextlib, datetime), sqlalchemy (text, AsyncSession, async_sessionmaker), and eventsource.observability for Tracer / create_tracer. No PostgreSQL driver is imported — asyncpg never appears — so the import succeeds on a base install regardless of which extras are present. See No optional-dependency guard.
Quick Reference
Constructor and full public surface at a glance. Import paths per Import Surface above.
PostgreSQLLockManager(
session_factory: async_sessionmaker[AsyncSession],
*,
holder_id: str | None = None,
tracer: Tracer | None = None,
enable_tracing: bool = True,
)
| Member | Signature | Returns | Raises | |
|---|---|---|---|---|
acquire |
`acquire(key: str, *, timeout: float \ | None = None, retry_interval: float = 0.1)` | async context manager yielding LockInfo |
LockAcquisitionError |
try_acquire |
await try_acquire(key: str) |
LockInfo on success, None if held elsewhere |
underlying driver exception on database error | |
release |
await release(key: str) |
None |
LockNotHeldError |
|
is_held |
await is_held(key: str) |
bool |
— | |
release_all |
await release_all() |
int (locks released) |
— | |
held_lock_count |
property (sync) | int |
— |
| Module-level name | Signature | Returns |
|---|---|---|
migration_lock_key |
migration_lock_key(tenant_id: UUID, operation: str = "migration") |
f"{operation}:{tenant_id}" |
LockInfo is a frozen dataclass with four fields: key: str, lock_id: int, acquired_at: datetime (UTC), holder_id: str | None.
Context-managed acquisition (preferred)
Blocking — waits indefinitely for the lock, releases on exit either way:
from eventsource.adapters.postgresql.locks import PostgreSQLLockManager
lock_manager = PostgreSQLLockManager(session_factory, holder_id="worker-1")
async with lock_manager.acquire("cutover:tenant-123") as lock_info:
print(lock_info.lock_id) # derived bigint, diagnostics only
await perform_cutover()
Bounded — polls pg_try_advisory_lock every retry_interval seconds until the deadline, then raises:
from eventsource.ports.exceptions import LockAcquisitionError
try:
async with lock_manager.acquire("cutover:tenant-123", timeout=5.0):
await perform_cutover()
except LockAcquisitionError as exc:
# exc.key, exc.reason ("Timeout after 5.0s"), exc.timeout
print(f"another instance holds {exc.key}: {exc.reason}")
Manual acquisition
try_acquire never waits and never raises on contention — it returns None. Release is the caller's responsibility, so pair it with try/finally:
lock_info = await lock_manager.try_acquire("cutover:tenant-123")
if lock_info is None:
return # someone else has it
try:
await perform_cutover()
finally:
await lock_manager.release("cutover:tenant-123")
Inspection and shutdown
await lock_manager.is_held("cutover:tenant-123") # this manager only
lock_manager.held_lock_count # sync property, no await
released = await lock_manager.release_all() # cleanup on shutdown
is_held, held_lock_count, and release_all report only locks held by this manager instance — they say nothing about other processes contending for the same key.
Migration key helper
from eventsource.ports.locks import migration_lock_key
key = migration_lock_key(tenant_id, "cutover") # "cutover:<uuid>"
async with lock_manager.acquire(key):
await perform_cutover()
Reminders
- PostgreSQL only — the manager issues
pg_advisory_lock,pg_try_advisory_lock, andpg_advisory_unlockas raw SQL with no fallback. - One open
AsyncSessionper held lock; N concurrent locks consume N pool connections. LockAcquisitionErrorandLockNotHeldErrorderive fromException, not from any eventsource error base.
`PostgreSQLLockManager`
class PostgreSQLLockManager:
def __init__(
self,
session_factory: async_sessionmaker[AsyncSession],
*,
holder_id: str | None = None,
tracer: Tracer | None = None,
enable_tracing: bool = True,
) -> None: ...
The manager owns a set of PostgreSQL advisory locks on behalf of one process. It is not a connection pool, a store, or a registry of locks held anywhere else — it is a bookkeeper for the locks it took, plus the sessions keeping them alive.
Instance state
Construction performs no I/O. It only records the collaborators and initializes empty bookkeeping:
| Attribute | Type | Role | |
|---|---|---|---|
_session_factory |
async_sessionmaker[AsyncSession] |
Checks out one session per lock acquisition | |
_holder_id |
`str \ | None` | Copied verbatim into every LockInfo.holder_id |
_tracer |
Tracer |
Wraps acquire and release in spans | |
_held_locks |
dict[str, tuple[AsyncSession, int]] |
Key → (session holding it, derived lock ID) | |
_lock |
asyncio.Lock |
Guards mutations of _held_locks |
_held_locks is the manager's whole notion of "held." is_held, release, release_all, and held_lock_count all read from it, so they answer a process-local question: did this manager object take this key and not yet give it back? A second PostgreSQLLockManager in the same process, or the same manager in another process, has a separate dictionary; none of these methods query PostgreSQL to find out who else is contending. Only an acquisition attempt does that.
Lifecycle of a lock
Every acquisition follows the same three steps, regardless of which entry point is used:
- Derive the
bigintlock ID from the string key (_key_to_lock_id, see Key-to-Lock-ID Derivation). - Check out a fresh
AsyncSessionfromsession_factoryand issuepg_advisory_lockorpg_try_advisory_lockon it. On failure the session is closed before the error propagates, so a failed attempt leaks nothing. - On success, record
key -> (session, lock_id)in_held_locksunder theasyncio.Lock, log atDEBUG, and hand back aLockInfo.
Release is the mirror image, funnelled through a single internal _release_lock: issue pg_advisory_unlock on the owning session, drop the key from _held_locks, and close the session. The unlock statement is wrapped in its own try/except that logs a WARNING rather than raising — a database error during release does not prevent the bookkeeping entry from being removed or the session from being closed, since both happen in a finally. The practical upshot is that release never raises for database reasons; the only exception it can raise is LockNotHeldError, and that comes from the public release() wrapper before _release_lock is reached.
Two acquisition styles
acquire() |
try_acquire() |
|
|---|---|---|
| Form | async context manager | plain coroutine |
| Waits? | yes (indefinitely, or up to timeout) |
never |
| Contention result | raises LockAcquisitionError (only when timeout set) |
returns None |
| Release | automatic, in finally |
caller must call release(key) |
| DB error | re-raised as LockAcquisitionError |
underlying driver exception propagates unchanged |
Prefer acquire(). Its finally block releases the lock on any exit path — normal return, exception, or cancellation of the enclosing task — so the lock cannot outlive the critical section. Reach for try_acquire() only when "someone else has it, move on" is a legitimate outcome and you cannot express the critical section as a with block.
Note the asymmetry in error handling between the two: acquire() translates database failures into LockAcquisitionError with a reason of "Database error: ...", while try_acquire() logs at ERROR and re-raises the driver's own exception. Code that must handle both uniformly should catch Exception around try_acquire(), not just LockAcquisitionError.
Thread- and task-safety
Mutations of _held_locks are serialized by an asyncio.Lock, which makes acquire, try_acquire, release, is_held, and release_all safe to call concurrently from multiple tasks in the same event loop. Two caveats:
held_lock_countis a plain synchronous property that readslen()without taking the guard. Its docstring flags it as not async-safe; treat it as a diagnostic reading, not a value to branch on in concurrent code.- The
asyncio.Lockis bound to the loop that created it. A manager is not shareable across event loops or OS threads; construct one per loop.
Re-acquiring a key the manager already holds is not guarded. PostgreSQL advisory locks are re-entrant within a session, but each acquisition here uses a new session, so a second acquire() of the same key from the same manager blocks against the first — a self-deadlock when timeout=None. The _held_locks entry is also overwritten by the second acquisition, orphaning the first session's bookkeeping. Do not nest acquisitions of the same key.
Typical construction
from sqlalchemy.ext.asyncio import async_sessionmaker, create_async_engine
from eventsource.adapters.postgresql.locks import PostgreSQLLockManager
engine = create_async_engine("postgresql+asyncpg://localhost/app", pool_size=10)
session_factory = async_sessionmaker(engine, expire_on_commit=False)
lock_manager = PostgreSQLLockManager(
session_factory,
holder_id=f"worker-{os.getpid()}",
)
A single manager instance is intended to be long-lived and shared across the tasks in a process — it is cheap to construct but holds the bookkeeping that makes release_all() meaningful at shutdown. Wire await lock_manager.release_all() into your shutdown path so locks are given back deliberately rather than relying on connection teardown.
Constructor
PostgreSQLLockManager(
session_factory: async_sessionmaker[AsyncSession],
*,
holder_id: str | None = None,
tracer: Tracer | None = None,
enable_tracing: bool = True,
) -> None
One positional parameter, three keyword-only. Only session_factory is required; the rest affect diagnostics rather than locking behaviour.
| Parameter | Type | Default | Effect | |
|---|---|---|---|---|
session_factory |
async_sessionmaker[AsyncSession] |
— | Source of the dedicated session opened per held lock | |
holder_id |
`str \ | None` | None |
Copied into LockInfo.holder_id; debugging label only |
tracer |
`Tracer \ | None` | None |
Explicit tracer; overrides enable_tracing when given |
enable_tracing |
bool |
True |
Whether to build an OpenTelemetry tracer when tracer is omitted |
__init__ performs no I/O and issues no SQL. It does not connect to PostgreSQL, does not validate that the engine behind session_factory speaks PostgreSQL, and does not verify that advisory-lock functions exist. A manager constructed against a SQLite engine builds successfully and fails only on the first acquire() or try_acquire(). Construction is therefore cheap and safe at import or startup time, before the database is reachable.
The body assigns five attributes and nothing else:
self._tracer = tracer or create_tracer(__name__, enable_tracing)
self._enable_tracing = self._tracer.enabled
self._session_factory = session_factory
self._holder_id = holder_id
self._held_locks: dict[str, tuple[AsyncSession, int]] = {}
self._lock = asyncio.Lock()
Two of those are worth noting for callers:
_held_locksstarts empty, so a freshly constructed manager reportsheld_lock_count == 0andis_held(key) is Falsefor every key — even when another process holds that key in PostgreSQL. Held-lock state is per instance and never rehydrated from the database._lockis anasyncio.Lockcreated eagerly. On modern Python it binds to the running loop on first use rather than at construction, but the manager is still single-loop: construct it inside the event loop that will use it, and do not share one instance across loops or threads.
There is no close(), __aenter__, or __aexit__ on the manager itself. The counterpart to construction is release_all(), which closes every session still held; see release_all().
Minimal construction
from sqlalchemy.ext.asyncio import async_sessionmaker, create_async_engine
from eventsource.adapters.postgresql.locks import PostgreSQLLockManager
engine = create_async_engine("postgresql+asyncpg://localhost/app")
session_factory = async_sessionmaker(engine, expire_on_commit=False)
lock_manager = PostgreSQLLockManager(session_factory)
Everything else is optional. The following two constructions behave identically with respect to locking; they differ only in what shows up in LockInfo and in your traces:
plain = PostgreSQLLockManager(session_factory)
labelled = PostgreSQLLockManager(
session_factory,
holder_id=f"worker-{os.getpid()}",
enable_tracing=False,
)
Keyword-only enforcement
holder_id, tracer, and enable_tracing sit behind a bare *, so they cannot be passed positionally:
PostgreSQLLockManager(session_factory, "worker-1") # TypeError
PostgreSQLLockManager(session_factory, holder_id="worker-1") # correct
How many managers to construct
One per process is the intended shape. The manager is stateful — its _held_locks dictionary is the only record of which sessions must be closed — so constructing a throwaway manager per operation defeats release_all() and makes is_held() answer about a scope narrower than you likely mean. Build one at application startup alongside the engine, share it across tasks, and release it during shutdown.
Constructing several managers against the same session_factory is legal and occasionally useful (independent holder_id labels, independent release_all() scopes), but each maintains its own bookkeeping, so two managers in the same process contend against each other for the same key exactly as two processes would — each acquisition opens its own PostgreSQL session, and advisory locks are session-scoped.
`session_factory`
session_factory: async_sessionmaker[AsyncSession]
The only required parameter, and the manager's sole route to the database. It is stored as self._session_factory and called — with no arguments — once per acquisition attempt:
session = self._session_factory()
That call happens in _acquire_lock (used by acquire()) and again in try_acquire(). Nothing else in the class touches the factory: there is no long-lived "control" session, no engine reference, and no reuse of a session across two different keys.
What the manager expects of it
| Expectation | Why |
|---|---|
| Callable with zero arguments | Invoked as self._session_factory() |
Produces an AsyncSession |
await session.execute(...) and await session.close() are called on the result |
| Bound to a PostgreSQL engine | The statements are pg_advisory_lock / pg_try_advisory_lock / pg_advisory_unlock |
| Backed by a pool with spare capacity | One session stays checked out for the whole time a lock is held |
The annotation is async_sessionmaker[AsyncSession], but the runtime requirement is only the first two rows — any zero-arg callable returning an object with async execute and close satisfies the code path, which is what lets unit tests substitute a mock factory. Type checkers will still demand the real async_sessionmaker, so production code should pass one.
The manager never calls commit(), rollback(), begin(), or flush() on the sessions it creates, and never uses them as async context managers. Advisory locks are session-scoped rather than transaction-scoped, so no transaction boundary is needed to make the lock durable for its lifetime. Session configuration that only affects ORM behaviour — expire_on_commit, autoflush, mapped-class registry — is therefore irrelevant here; the factory's only meaningful attribute is the engine it is bound to.
It is not validated at construction
__init__ stores the factory and returns. It does not call it, does not open a connection, and does not inspect the engine's dialect. A manager built over a SQLite or MySQL engine constructs cleanly and fails on first use — see PostgreSQL-Only Constraint for the shape of that failure, which differs between acquire() (LockAcquisitionError with reason="Database error: ...") and try_acquire() (driver exception re-raised after an ERROR log).
Session lifetime is lock lifetime
Each acquisition holds its session open for exactly as long as it holds the lock:
session = self._session_factory()— a connection is checked out of the pool.SELECT pg_advisory_lock(...)(or thepg_try_advisory_lockpoll loop) runs on that session.- On success the session is stored in
_held_locks[key]and stays open, idle, holding the lock. _release_lockissuespg_advisory_unlockon the same session, thenawait session.close()in afinally, returning the connection to the pool.
Failure paths close the session before propagating: the timeout branch closes it before raising LockAcquisitionError, the generic except Exception branch closes it before re-raising as LockAcquisitionError, and try_acquire() closes it both when the lock is already held elsewhere (returning None) and when the statement errors. A failed acquisition therefore does not leak a connection.
The direct consequence for sizing: the pool must accommodate one connection per concurrently held lock, on top of everything else the application is doing. Sharing an application's main session_factory with the lock manager is fine as long as that headroom exists; if locks are held for long stretches or in large numbers, give the manager a factory over its own small engine so a stuck lock cannot starve request handling. See Session and Connection Pool Model for the full treatment.
Typical wiring
from sqlalchemy.ext.asyncio import async_sessionmaker, create_async_engine
from eventsource.adapters.postgresql.locks import PostgreSQLLockManager
engine = create_async_engine(
"postgresql+asyncpg://localhost/app",
pool_size=10,
max_overflow=5,
)
session_factory = async_sessionmaker(engine, expire_on_commit=False)
lock_manager = PostgreSQLLockManager(session_factory)
A dedicated engine, when lock traffic should be isolated from application traffic:
lock_engine = create_async_engine(
"postgresql+asyncpg://localhost/app",
pool_size=4, # ceiling on simultaneously held locks
max_overflow=0,
)
lock_manager = PostgreSQLLockManager(
async_sessionmaker(lock_engine, expire_on_commit=False),
holder_id=f"worker-{os.getpid()}",
)
Note that a pool ceiling becomes a lock ceiling: with pool_size=4 and max_overflow=0, a fifth concurrent acquisition blocks waiting for a connection before it ever reaches pg_advisory_lock, and that wait is governed by the engine's pool_timeout, not by the timeout argument to acquire().
Sharing with the rest of the application
The factory may be the same one used by other session-based PostgreSQL-backed components — the manager holds no exclusive claim on it and adds no state to the sessions it borrows. (Note that PostgreSQLEventStore itself is constructed from an AsyncEngine, not a session factory, so it is not a sharing candidate here — but PostgreSQLSnapshotStore and other session-factory-based components are.) What it must not be given is a single already-open AsyncSession wrapped in a lambda: every acquisition needs its own session, because two locks sharing one PostgreSQL session would also share lock ownership, and closing one would drop the other.
Connection reuse after a release failure
_release_lock catches any error from pg_advisory_unlock, logs a WARNING, and proceeds to close the session regardless. Closing returns the connection to the pool, and pool checkin rolls back the transaction — but a rollback does not clear session-level advisory locks. In the rare case where the unlock statement itself fails yet the connection survives and is recycled, the advisory lock can outlive the manager's bookkeeping, since it is released only when that backend connection actually terminates. Treat repeated "Error releasing advisory lock" warnings as a signal to recycle the engine's pool (await engine.dispose()) rather than as noise.