Postgres/Redshift adapters mask SSL syscall error with "connection already closed"

[cmcarthur]

Issue

Issue description

A recent Sinter run of dbt archive encountered the dreaded:

SSL SYSCALL error: EOF detected

Relevant part of the call stack:

  File "/usr/local/lib/python3.5/dist-packages/dbt/runner.py", line 89, in call_runner
    result = runner.safe_run(manifest)
  File "/usr/local/lib/python3.5/dist-packages/dbt/node_runners.py", line 121, in safe_run
    self.adapter.release_connection(self.profile, node_name)
  File "/usr/local/lib/python3.5/dist-packages/dbt/adapters/default/impl.py", line 494, in release_connection
    cls.rollback(to_release)
  File "/usr/local/lib/python3.5/dist-packages/dbt/adapters/default/impl.py", line 613, in rollback
    connection.get('handle').rollback()

When it tries to rollback the transaction on that connection handle, postgres raises another exception which masks the original one and prints out connection already closed.

There was an earlier Failed to release connection! error -- I don't know if the two are related.

System information

dbt version: 0.11.0
platform: Sinter
python version: 3.5

Steps to reproduce

Difficult to repro since it depends on Postgres server failing in a specific way. However I imagine manually killing a postgres connection handle in the middle of an archive job would create similar behavior.

[beckjake]

Fixed (hopefully) in #1000