An insert-only Python client for River packaged in the riverqueue
package on PyPI. Allows jobs to be inserted in Python and run by a Go worker, but doesn't support working jobs in Python.
Your project should bundle the riverqueue
package in its dependencies. How to go about this will depend on your toolchain, but for example in Rye, it'd look like:
rye add riverqueue
Initialize a client with:
import riverqueue
from riverqueue.driver import riversqlalchemy
engine = sqlalchemy.create_engine("postgresql://...")
client = riverqueue.Client(riversqlalchemy.Driver(engine))
Define a job and insert it:
@dataclass
class SortArgs:
strings: list[str]
kind: str = "sort"
def to_json(self) -> str:
return json.dumps({"strings": self.strings})
insert_res = client.insert(
SortArgs(strings=["whale", "tiger", "bear"]),
)
insert_res.job # inserted job row
Job args should comply with the riverqueue.JobArgs
protocol:
class JobArgs(Protocol):
kind: str
def to_json(self) -> str:
pass
kind
is a unique string that identifies them the job in the database, and which a Go worker will recognize.to_json()
defines how the job will serialize to JSON, which of course will have to be parseable as an object in Go.
They may also respond to insert_opts()
with an instance of InsertOpts
to define insertion options that'll be used for all jobs of the kind.
We recommend using dataclasses
for job args since they should ideally be minimal sets of primitive properties with little other embellishment, and dataclasses
provide a succinct way of accomplishing this.
Inserts take an insert_opts
parameter to customize features of the inserted job:
insert_res = client.insert(
SortArgs(strings=["whale", "tiger", "bear"]),
insert_opts=riverqueue.InsertOpts(
max_attempts=17,
priority=3,
queue="my_queue",
tags=["custom"]
),
)
Unique jobs are supported through InsertOpts.unique_opts()
, and can be made unique by args, period, queue, and state. If a job matching unique properties is found on insert, the insert is skipped and the existing job returned.
insert_res = client.insert(
SortArgs(strings=["whale", "tiger", "bear"]),
insert_opts=riverqueue.InsertOpts(
unique_opts=riverqueue.UniqueOpts(
by_args=True,
by_period=15*60,
by_queue=True,
by_state=[riverqueue.JobState.AVAILABLE]
)
),
)
# contains either a newly inserted job, or an existing one if insertion was skipped
insert_res.job
# true if insertion was skipped
insert_res.unique_skipped_as_duplicated
Unique job insertion takes a Postgres advisory lock to make sure that its uniqueness check still works even if two conflicting insert operations are occurring in parallel. Postgres advisory locks share a global 64-bit namespace, which is a large enough space that it's unlikely for two advisory locks to ever conflict, but to guarantee that River's advisory locks never interfere with an application's, River can be configured with a 32-bit advisory lock prefix which it will use for all its locks:
client = riverqueue.Client(riversqlalchemy.Driver(engine), advisory_lock_prefix=123456)
Doing so has the downside of leaving only 32 bits for River's locks (64 bits total - 32-bit prefix), making them somewhat more likely to conflict with each other.
Use #insert_many
to bulk insert jobs as a single operation for improved efficiency:
num_inserted = client.insert_many([
SimpleArgs(job_num=1),
SimpleArgs(job_num=2)
])
Or with InsertManyParams
, which may include insertion options:
num_inserted = client.insert_many([
InsertManyParams(args=SimpleArgs(job_num=1), insert_opts=riverqueue.InsertOpts(max_attempts=5)),
InsertManyParams(args=SimpleArgs(job_num=2), insert_opts=riverqueue.InsertOpts(queue="high_priority"))
])
To insert jobs in a transaction, open one in your driver, and pass it as the first argument to insert_tx()
or insert_many_tx()
:
with engine.begin() as session:
insert_res = client.insert_tx(
session,
SortArgs(strings=["whale", "tiger", "bear"]),
)
The package supports River's asyncio
(asynchronous I/O) through an alternate AsyncClient
and riversqlalchemy.AsyncDriver
. You'll need to make sure to use SQLAlchemy's alternative async engine and an asynchronous Postgres driver like asyncpg
, but otherwise usage looks very similar to use without async:
engine = sqlalchemy.ext.asyncio.create_async_engine("postgresql+asyncpg://...")
client = riverqueue.AsyncClient(riversqlalchemy.AsyncDriver(engine))
insert_res = await client.insert(
SortArgs(strings=["whale", "tiger", "bear"]),
)
With a transaction:
async with engine.begin() as session:
insert_res = await client.insert_tx(
session,
SortArgs(strings=["whale", "tiger", "bear"]),
)
The package exports a py.typed
file to indicate that it's typed, so you should be able to use MyPy to include it in static analysis.
Our read is that SQLAlchemy is the dominant ORM in the Python ecosystem, so it's the only driver available for River. Under the hood of SQLAlchemy, projects will also need a Postgres driver like psycopg2
or asyncpg
(for async).
River's driver system should enable integration with other ORMs, so let us know if there's a good reason you need one, and we'll consider it.
See development.