when the ds becomes too large to store it all in memory, is tying with MLDBM the preferred paradigm? What are the alternatives?

It really depends upon what is in the data structure and how you are using it. Some of the questions that might influence the best choice are:

• How long does the application(s) that use it run for?

  A web app that runs hundreds or thousands of times an hour for a few milliseconds each time might require a different solution to a data processing app that loads once a day or week and runs for minutes or hours each time.

• How much of the data structure is used in each run?

  If only one or two rows are used per run, it might make more sense to leave the data on disk and structure the file such that it can be randomly accessed.

• If a large proportion of the data is used for each run, is the structure traverse once or multiple times? Is it traversed serially or randomly?
• Is the data structure read-only, occasionally written or frequently written?
• Does only one application use the data concurrently? Or many?
• Can the processing of the large data set be encoded using SQL to reduce it to producing small results sets?

The best solution may depend upon the answers to some or all of these questions, and more that arise from those answers. A clear description of the data set and how it is used would be the quickest way of eliciting good answers.