Best practices page for documentation

[pavlis]

Working on building a production workflow for a full up test of the usarray data set made me realize we need to create one or more "best practices" or "practical guildelines" pages. I am thinking the way it should evolve is this:

1. First, if there is agreement I can start a wiki page on the topic. That seems like the best start as multiple people can easily contribute to items in such a wiki page. If it grows too bit it could be easily split into multiple categories - if we are so lucky.
2. We need a section of the user manual with a similar title that would be more formal and tutorial.

The other issue is what topics do we need? Here are some suggestions. I urge the rest of the development team to append to this list:

1. Practical issues about running docker on a desktop. I'm running into issues at the moment with the jupyter notebook crashing leaving no hint why in the log files. I am pretty sure the container is running out of memory. We need to preserve the hint that when the notebook crashes the most common reason seems to be memory faults.
2. There are general, practical issues about memory management. This could be several topics actually.\
3. There are issues about how to structure a workflow. e.g. to manage memory it is often critical to do time windowing as early as possible in a set of map calls.
4. A special case of 3 is something we need to probably work on. I suspect strongly that performance can be improved if a set of map, map, map, map, ..., map lines are replaced by small wrapper functions that do all the stuff in one function passed to map. That would likely reduce serialization overhead, but that is a hypothesis only.
5. There are mulitple possible topics on io performance.
6. There are important practical issues in how to debug a workflow running inside jupyter. A forum to put simple to more complex solutions would be helpful to all and any users I'm sure.
7. There is a basic practical thing about benchmarking a workflow on a smaller data set before trying to run in on a very large data set on a cluster. I don't think we have very good advice to give on this just yet, but we can do something useful. I think the basic advice is get your workflow running on a desktop with a small data set. Then make an assumption about scaling (there are lots of caveats there), use our paper as a guideline for how much you can buy from multiprocessing, and compute the time you would need to run the large data sets.
8. Advice on checkpointing large runs an starting in the middle if the job dies.
9. Ensemble or not to ensemble.

A start. Please add to this list and discussion.