Deprecated thread local strategy

[greenape]

We're stuck on an old version of pandas when using flowmachine, because new pandas uses new sqlalchemy which deprecated threadlocal strategy.

We assume threadlocal throughout flowmachine's codebase at present in a way which is quite challenging to unpick, because one of the side effects of it is that one can be in a transaction in a very concealed way. For example, everything in write_query_to_cache is in a single transaction, but this is not immediately obvious.

We're quite inconsistent about when we pass in a connection (and what a connection is when passed), and when we summon one from context using get_db(), which also makes this a bit difficult to resolve.

[greenape]

I'm wondering if the way to resolve this is to never pass the db connection?

The db connection is available from context, what if we extend that and say db accessible only through context handler, there are two handlers: fresh transaction and nested within current transaction?

[greenape]

Although see https://docs.sqlalchemy.org/en/14/core/connections.html#arbitrary-transaction-nesting-as-an-antipattern for a counter to that, suggesting we should be explicitly passing around instead, which tbh is probably better, but more annoying to migrate to given where we're at.