Hotfix/add connection pool for postgresql only

[Quinn-Fang]

Add connection pool and make it work correctly.

Description

So basically, I am using Tortoise in my Project which needs to handle more than 20000 QPS, thus the asyncpg connection pool is required. So I've tried to incorporate PR 141's asyncpg/client and base/client into tortoise, so at lease I can get a temporary connection pool version of tortoise to get my project work. But after a little pressure test, (somewhere around 120 requests/s, 3 create per transaction, 40 transactions/s), error occurs, says a "transaction is already closed", which means the ddl s inside the same transaction were using different connections. it's because the execute_insert function inside AsyncpgDBClient is creating new connections wether it's inside a transaction or not, so in order to let all ddl s use same connection, I use a contextvars to store the transactionwrapper instance in the current coroutine context.

[grigi]

Interesting work. I'm glad you managed to work around that same issue I got when I was looking at it months ago.

So is it only execute_insert that forced a new connection? I'm pretty sure that all the execute_* ones do… (The self.acquire_connection(…) does it?)
I see you added two more contextvars, current_transaction (global) and self._current_transaction? Why is that?

[grigi]

I spent some time looking at this and I think I see what you did:

The transaction wrapper check:

transaction_wrapper = current_transaction.get()
if transaction_wrapper:
    transaction_wrapper._connection

is logic that should really be run by:

async with self.acquire_connection() as connection:

Also, you used your own connection-local connection wrapper object, instead of the global one.

🤔

[grigi]

Hi @Quinn-Fang We fixed the context leak, and have PR #229 just about done right now that should implement pooling.
Thank you for this PR, it was very helpful in helping to find the root cause.

Could you please have a look at & test #229 ?

[grigi]

Closing this as we have a pooling implementation that seemingly works in master (but not yet released)