Pandas is mostly columnar in nature. If you're heavy into Python idioms it's fairly easy to grok, imo, but there is a subtle learning curve about what is in and out of kernel. Since in-kernel operations are about 50-100x faster, it's obvious when you did the wrong thing, but it doesn't show until data sets are huge.

I tried doing something regarding browser hits from Akamai data in Pandas and 3 different SQL databases (mysql, postgres, sqlite) and nothing came close to pandas for holding 150m hits (one day's worth across our properties) in memory as well as Pandas. Especially with Dask Dataframes mixed in. No competition for the effort involved.

Using tools such as sqlmagic or pyodbc for ipython (below) you can let the sql database perform most of the heavy lifting of getting "raw" result sets into a dataframe and then perform some lighter, possibly more ptyhon-idiomatic tweaks to the data once the focused set is loaded in-memory.

  def getdataframe(sql):
     try:
          connection = pyodbc.connect("DSN=myDSN", autocommit=True)
          return pandas.read_sql(sql, connection)
      except Exception as e:
          print repr(e)
      finally:
          connection.close()