[CT-2377] [Spike] Performance improvement to memory footprint

[nathaniel-may]

When a macro runs a call statement with fetch_result set to True, that result may be held in memory longer than it needs to be.

[jtcohen6]

@nathaniel-may Thanks for opening! (I'd been meaning to)

More context in this internal Slack thread

As a general rule, should we be garbage-collecting the context from each NodeRunner, after that node finishes compiling/executing?

[iknox-fa]

Spike resulted in a fix since it was easy-- but for future reference:

The issue at hand seems to be that ProviderContext and it's subclasses have an attribute dict called sql_results where all of the results from statement calls get stored via the store_results and/or store_raw_results context functions. Over time the number of entries in that dict grows and so does the memory used. The ideal solution would be to clear the dict when each calling macro is complete but unfortunately, determining the lifecycle of these classes once instantiated is a little tricky because it happens in the depths of the jinja code where we can't really effect the outcome (I think... I am def getting into language team things here).

That said, we can control what happens when we read the values via the load_results context function, so we can do something clever and remove the value from the dict as it's being returned. What this means is that <type>Context.sql_results will only ever get as big as the data you've put in but not yet assigned to a variable in jinja. For most use cases I think this works, however it does mean that we can only ever call {% set foo = load_result(result_name) %} once.

Instinctually, that feels like a best practice anyway, but I'd be curious to hear what others think.