Simple query that works with sqlalchemy fails with databases

[trvrmcs]

A simple COUNT(*) works using sqlalchemy.Engine.execute(), but fails using databases.Database.fetch_all(), raising an IndeterminateDatatypeError exception.

import asyncio
import databases
import sqlalchemy
from sqlalchemy import Column, MetaData, Text, Table, select, func


url = "postgresql://test:password@localhost/example"
engine = sqlalchemy.create_engine(url)


metadata = MetaData()
People = Table("people", metadata, Column("name", Text()))
QUERY = select([func.count("*")]).select_from(People)


def setup():
    engine.execute("CREATE TABLE IF NOT EXISTS people(name text)")


def test_sync():
    "this works"
    engine.execute(QUERY).fetchall()


async def test_async():
    "this fails with 'asyncpg.exceptions.IndeterminateDatatypeError: could not determine data type of parameter $1'"
    database = databases.Database(url)
    await database.connect()
    await database.fetch_all(QUERY)
    await database.disconnect()


if __name__ == "__main__":
    setup()
    test_sync()

    asyncio.get_event_loop().run_until_complete(test_async())

[tomchristie]

Steps to progress this issue would include:

• What's the equivalent raw string query here, and does that succeed or fail?
• Does this success or fail on mysql, sqlite?

The exception itself is from the asyncpg library, so it's so either an issue there, or it's to do with the query compilation in databases.

[gvbgduh]

It looks like it's due to a subtle difference in the underlying driver.
await database.fetch_all(QUERY) will render the query and perform the command as

raw_connection.fetch('select count($1) from notes;', '*')

that will raise an exception at the asyncpg level

asyncpg.exceptions.IndeterminateDatatypeError: could not determine data type of parameter $1

and to resolve the confusion asyncpg expects it as

raw_connection.fetch('select count($1::text) from notes;', '*')

Another way to resolve it would be to define the query as
select([func.count()]).select_from(notes)
that'll be rendered as

'SELECT count(*) AS count_1 \nFROM notes'

that doesn't confuse asyncpg.

For engine.execute(QUERY).fetchall() the underlying driver is psycopg2, that receives the query as

engine.execute('select count(%s) from notes;', '*').fetchall()

that it can execute.

[tomchristie]

Is this resolvable for us? Is there anything we can set in the dialect that'd ensure the query gets compiled to 'select count($1::text) from notes;'?

[gvbgduh]

@tomchristie I'm not sure, it's getting tricky as sqlalchemy "supports" psycopg2 (it'll cover aiopg well tho) as the postgres driver and has no support for asyncpg. Those two have subtle differences in general and in the query syntax details particularly and the sqlalchemy's query render isn't aware of it.

In this particular case 'select count(%s) from notes;', '*' would be the best as it suits both drivers.

It's a quite confusing point to me as I'm afraid it might get pretty complicated to control the query rendering might require to go quite deep in adjusting the sqlalchemy logic, where it might have more sense to do at the ORM level. However, on the other hand, we have integration with sqlalchemy core already.

[tomchristie]

May be worth looking into if asyncpgsa has had any similar issues raised, and if they've resolved them or not.

[waydegilliam]

Any updates on this issue?

[waydegilliam]

Another solution: compile the query as a string with your database's dialect and use the "literal_binds" kwarg when compiling. This substitutes the parameters into raw SQL which can be dangerous if the parameters are from an untrusted input.

Here's an example for PostgreSQL:

from sqlalchemy.dialects import postgresql

# other imports, database connection, etc. ...

C_QUERY = str(QUERY.compile(compile_kwargs={"literal_binds": True}, dialect=postgresql.dialect()))
res = await database.fetch_all(C_QUERY)

Annoying since you'd have to do this for every query, but I'm using a custom subclass of databases.Database that handles this automatically.