Question

GinoEngine is the core of GINO. It acts like a pool of connections but also does the work of assembling everyone together:

Under the hood, engine is associated with a specific dialect instance on creation, e.g. asyncpg dialect. The dialect is actually a set of classes that implements GINO dialect API, offering all the details about how to operate on this specific database. In the diagram, gray color means internal, while green means touchable by end users.

During creation, the engine will also ask the dialect to create a database connection pool for it. The pool type is also a part of the dialect API, because asynchronous database drivers usually have their own pool implementation, thus their GINO dialects should hide such implementation differences behind the unified diagram API for engine to use.

注解

In SQLAlchemy, database drivers are supposed to follow the DB-API standard, which does not usually provide a pool implementation. Therefore, SQLAlchemy has its own pool implementation, created directly in engine. This is where this diagram doesn't fit SQLAlchemy.

The pool creates raw connections, not the GinoConnection green in the diagram. The connection in the diagram is a many-to-one wrapper of the raw connection, because of the reuse and lazy features, we'll get to that part later. The connection is created by the engine, thus inherits the same dialect, and is used for running queries.

On the outer side, SQLAlchemy queries can be executed directly on the engine or connection. When on engine, it will try to acquire a reusable connection to actually execute the connection, and release the connection after use.

注解

Another difference to SQLAlchemy here: GINO execution methods always return final results, while in SQLAlchemy accessing the result may cause further implicit database accesses. Therefore GINO engine immediately releases the connection when the execution method on the engine returns, but SQLAlchemy can only release the connection implicitly when the result data is found exhausted.

By immediately releasing a connection, GINO may not release the related raw connection when the raw connection was reused from another parent connection. We'll get to this later.

GINO also supports implicit execution without having to specify an engine or connection explicitly. This is done by binding the engine to the db instance, also known as the MetaData or the Gino instance. You may possibly bind a GinoConnection instance, but that is greatly not recommended because it is very much untested.

At last, as the ORM / CRUD feature, models are just add-ons on top of everything else to generate queries. The parent model class is connected to a db instance on creation, therefore the models can do implicit execution too if their db has a bind.

Then let's get to some details.

Creating Engines

GINO reuses the strategy system SQLAlchemy provides to create engines. The name of GINO's strategy to create asynchronous GinoEngine is just gino, but only available after gino is imported:

import gino, sqlalchemy

async def main():
    e = await sqlalchemy.create_engine('postgresql://...', strategy='gino')
    # e is a GinoEngine

小技巧

Please read this SQLAlchemy document to learn about writing database URLs.

Also the GINO strategy replaces the default driver of dialect postgresql:// from psycopg2 to asyncpg, so that you don't have to replace the URL as it may be shared between GINO and vanilla SQLAlchemy in parallel. Alternatively, you can explicitly specify the driver to use by postgresql+asyncpg://... or just asyncpg://....

GINO also offers a shortcut as gino.create_engine(), which only sets the default strategy to gino and does nothing more. So here is an identical example:

import gino

async def main():
    e = await gino.create_engine('postgresql://...')
    # e is also a GinoEngine

As you may have noticed, when using the GINO strategy, create_engine() returns a coroutine, which must be awaited for result. Because it will create a database connection pool behind the scene, and actually making a few initial connections by default.

For it is just SQLAlchemy create_engine(), the same rules of parameters apply in GINO too. Well for now, GINO only supports a small amount of all the parameters listed in SQLAlchemy document (we are working on it!):

For Dialect:

• isolation_level

• paramstyle

For Engine:

• echo

• execution_options

• logging_name

While these parameters are discarded by GINO:

• module

In addition, keyword arguments for creating the underlying pool is accepted here. In the case of asyncpg, they are from create_pool(). For example, we can create an engine without initial connections:

e = await gino.create_engine('postgresql://...', min_size=0)

Similar to SQLAlchemy, GINO also provides shortcut to create engine while setting it as a bind. In SQLAlchemy it is like this:

import sqlalchemy

metadata = sqlalchemy.MetaData()
metadata.bind = 'postgresql://...'

# or in short

metadata = sqlalchemy.MetaData('postgresql://...')

This implicitly calls create_engine() under the hood. However in GINO, creating an engine requires await, it can no longer be hidden behind a normal assignment statement. Therefore, GINO removed the assignment magic in subclass Gino, reverted it to simple assignment:

import gino

db = gino.Gino()

async def main():
    # db.bind = 'postgresql://...' doesn't work!! It sets a string on bind
    engine = await gino.create_engine('postgresql://...')
    db.bind = engine

And provided a shortcut to do so:

engine = await db.set_bind('postgresql://...')

And another simpler shortcut for one-time usage:

db = await gino.Gino('postgresql://...')

To unset a bind and close the engine:

engine, db.bind = db.bind, None
await engine.close()

Or with a shortcut correspondingly:

await engine.pop_bind().close()

Furthermore, the two steps can be combined into one shortcut with asynchronous context manager:

async with db.with_bind('postgresql://...') as engine:
    # your code here

Managing Connections

With a GinoEngine at hand, you can acquire connections from the pool now:

conn = await engine.acquire()

Don't forget to release it after use:

await conn.release()

Yes this can be easily missing. The recommended way is to use the asynchronous context manager:

async with engine.acquire() as conn:
    # play with the connection

Here conn is a GinoConnection instance. As mentioned previously, GinoConnection is mapped to an underlying raw connection, as shown in following diagram:

Each column has at most one actual raw connection, and the number is the sequence the connections are created in this example. It is designed this way so that GINO could offer two features for connection management: reuse and lazy. They are keyword arguments on acquire() and by default switched off.

reuse

When acquiring a GinoConnection (2), GINO will borrow a raw connection (1) from the underlying pool first, and assign it to this GinoConnection (2). This is the default behavior of acquire() with no arguments given. Even when you are nesting two acquires, you still get two actual raw connection borrowed:

async with engine.acquire() as conn1:
    async with engine.acquire() as conn2:
        # conn2 is a completely different connection than conn1

But sometimes conn2 may exist in a different method:

async def outer():
    async with engine.acquire() as conn1:
        await inner()

async def inner():
    async with engine.acquire() as conn2:
        # ...

And we probably wish inner could reuse the same raw connection in outer to save some resource, or borrow a new one if inner is individually called without outer:

async def outer():
    async with engine.acquire() as conn1:
        await inner(conn1)

async def inner(conn2=None):
    if conn2 is None:
        async with engine.acquire() as conn2:
            # ...
    else:
        # the same ... again

This is exactly the scenario reuse could be useful. We can simply tell the acquire() to reuse the most recent reusable connection in current context by setting reuse=True, as presented in this identical example:

async def outer():
    async with engine.acquire() as conn1:
        await inner(conn1)

async def inner():
    async with engine.acquire(reuse=True) as conn2:
        # ...

Back to previous diagram, the blue GinoConnection instances (3, 4, 6) are "reusing connections" acquired with reuse=True, while the green ones (2, 5, 7) are not, thus they become "reusable connections". The green reusable connections are put in a stack in current context, so that acquire(reuse=True) always reuses the most recent connection at the top of the stack. For example, (3) and (4) reuse the only available (2) at that moment, therefore (2, 3, 4) all map to the same raw connection (1). Then after (5), (6) no longer reuses (2) because (5) is now the new head of the stack.

小技巧

By context, we are actually referring to the context concept in contextvars the new module in Python 3.7, and its partial backport aiocontextvars. Simply speaking, you may treat a series of function calls in a chain as in the same context, even if there is an await. It's something like a thread local in asyncio.

GinoConnection (2) may be created through acquire(reuse=True) too - because the stack is empty before (2), there is nothing to reuse, so (2) upgraded itself to a reusable connection.

Releasing a reusing connection won't cause the reused raw connection being returned to the pool, only directly releasing the reused GinoConnection can do so. Connections should be released in the reversed order as they are acquired, but if the reused connection is released before reusing connections by accident, then all the reusing connections depending on it will turn closed because they are reusing the same raw connection which is returned to the pool, any further execution will fail. For example, if (3) is released first, then (2) and (4) are still functional. But if (2) is released first, then (3) and (4) will be released implicitly and are no longer usable any more.

lazy

As you may have found, GinoConnection (5) does not have an underlying raw connection, even when it is reused by (6). This is because both (5) and (6) set lazy=True on acquire.

A lazy connection will not borrow a raw connection on creation, it will only do so when have to, e.g. when executing a query or starting a transaction. For example, GinoConnection (7) is acquired lazily without a raw connection, and (8) is only created when a query is executed on (7):

async with engine.acquire(lazy=True) as conn:  # (7)
    await conn.scalar('select now()')          # (8)

On implementation level, lazy is extremely easy in acquire(): if lazy=False then borrow a raw connection, else do nothing. That's it. Before executing a query or starting a transaction, GinoConnection will always try to borrow a raw connection if there is none present. This allows GINO to "transiently release" a raw connection, while all GinoConnection mapped to this raw connection are put in lazy mode (again). This is especially useful before you need to run some networking tasks in a database-related context - the networking task may take a long time to finish, we don't want to waste a connection resource checked out for nothing. For example:

async with engine.acquire(lazy=True) as conn:  # (7)
    await conn.scalar('select now()')          # (8)
    await conn.release(permanent=False)        # release (8)
    await asyncio.sleep(10)                    # simulate long I/O work
    await conn.scalar('select now()')          # re-acquire a new raw connection,
                                               #   not necessarily the same (8)

When used together with reuse, at most one raw connection may be borrowed for one reusing chain. For example, executing queries on both (5) and (6) will result only one raw connection checked out, no matter which executes first. It is also worth noting that, if we set lazy=False on (6), then the raw connection will be immediately borrowed on acquire, and shared between both (5) and (6). It's been quite a while, let me post the same diagram again:

reusable

Usually, you don't have to worry about the two options reuse and lazy, using the default acquire() will always create a concrete GinoConnection with a new raw connection with it. It is only that they are by default reusable (the green ones). If you need an absolutely isolated unique connection that has no risk being reused, you may use reusable=False on acquire. As shown in the diagram, the unreusable GinoConnection is an orphan away from any stack:

async with engine.acquire():                    # (2)
    async with engine.acquire(reusable=False):  # the unreusable connection
        async with engine.acquire(reuse=True):  # (3)

Unreusable connections can be lazy. But it is usually meaningless to specify both reuse=True and reusable=False at the same time, because reusing connections are always unusable - they are also not in the stack. You cannot reuse a reusing connection, you only reuse a reusable connection in the stack. Making a reusing connection unreusable doesn't make its related reusable connection unreusable. Hmm if this is getting more confusing, just don't use acquire(reuse=True, reusable=False) unless you know what it does.

current_connection

Except for all scenarios supported by above three options, there is still one left out: we may want to acquire a reusing-only connection. There is no such option to do so, but GINO could do the same thing through current_connection which is always the reusable GinoConnection at the top of current stack, or None if current stack is empty.

小技巧

The different between current_connection and acquire(reuse=True) is, the latter always produces a GinoConnection, while the former may not.

Executing Queries

Once you have a GinoConnection instance, you can start executing queries with it. There are 6 variants of the execute method: all(), first(), one(), one_or_none(), scalar() and status(). They are basically the same: accepting the same parameters, calling the same underlying methods. The difference is how they treat the results:

• all() returns all results in a list, which may be empty when the query has no result, empty but still a list.

• first() returns the first result directly, or None if there is no result at all. There is usually some optimization behind the scene to efficiently get only the first result, instead of loading the full result set into memory.

• one() returns exactly one result. If there is no result at all or if there are multiple results, an exception is raised.

• one_or_none() is similar to one(), but it returns None if there is no result instead or raising an exception.

• scalar() is similar to first(), it returns the first value of the first result. Quite convenient to just retrieve a scalar value from database, like NOW(), MAX(), COUNT() or whatever generates a single value. None is also returned when there is no result, it is up to you how to distinguish no result and the first value is NULL.

• status() executes the query and discard all the query results at all. Instead it returns the execution status line as it is, usually a textual string. Note, there may be no optimization to only return the status without loading the results, so make your query generate nothing if you don't want any result.

By "result", I meant RowProxy of SQLAlchemy - an immutable row instance with both tuple and dict interfaces. Database values are translated twice before they are eventually stored in a RowProxy: first by the database driver (dialect) from network payload to Python objects (see Type Conversion of how asyncpg does this), second by SQLAlchemy result_processor() depending on the actual type and dialect.

The arguments taken by these 4 methods are identical to the ones accepted by SQLAlchemy execute() (click to read more), usually a plain string of SQL directly or a SQLAlchemy query clause, followed by query parameters. In the case when multiple dictionaries are given to multiparams, all 4 methods will always return None discarding all results. Likewise, the parameter values are processed twice too: first by bind_processor() then the database driver.

GINO also supports SQLAlchemy execution_options() provided either on engine level, connection level or on queries. At the moment we are working on being compatible with SQLAlchemy execution options. In the mean while, GINO provides several new execution options, for example enabling return_model and providing a model will make all() and first() return ORM model instance(s) instead of RowProxy instance(s). See also execution_options() for more information.

In addition, GINO has an iterate() method to traverse the query results progressively, instead of loading all the results at once. This method takes the same arguments as the other 4 execute methods do, and follows the same rule of data handling. For now with asyncpg, this creates a server-side cursor.

Implicit Execution

Acquire a GinoConnection and execute queries on it, that is the most explicit way. You can also execute queries on a GinoEngine instance. In this case, a connection will be acquired with reuse=True for you implicitly, and released after returning:

await engine.scalar('select now()')

Equals to:

async with engine.acquire(reuse=True) as conn:
    await conn.scalar('select now()')

This allows you to easily write connectionless code. For example:

async def get_now():
    return await engine.scalar('select now()')

async def main():
    async with engine.acquire():
        now = await get_now()
        await engine.status('UPDATE ...')

In this example, main() will take only one raw connection. get_now() can also work alone out of any acquire() context, thanks to reuse.

Furthermore, GINO provides the same query APIs on Gino directly. They are simply delegates to corresponding API methods on the bind. This allows even engine-less programming:

db = gino.Gino()

async def get_now():
    return await db.scalar('select now()')

async def main():
    async with db.with_bind('postgresql://...'):
        now = await get_now()
        await db.status('UPDATE ...')

注解

In this example we didn't put the two queries in an acquire() block, so they might be executed in two different connections.

At last, the SQLAlchemy implicit execution on queries also work in GINO, under an extension named gino:

await users_table.select().gino.all()

By default, the extension GinoExecutor is injected on Executable as a property of name gino at the creation of Gino instance. Therefore, any Executable object has the gino property for implicit execution. Similarly, the execution methods calls the corresponding ones on the bind of the db instance.