C[omp]ute

From: andrew cooke <andrew@...>

Date: Tue, 2 Aug 2011 19:57:31 -0400

This is a brief summary of the things I've learnt while using Google's
AppEngine Data Store - a "NoSQL" database designed for high performance.


  1 - Do this!  I was wary of AppEngine because of lock-in, etc, but you can
      easily get Django working, which avoids learning a whole new framework,
      and Django non-rel has the promise to liberate you completely, if needed
      (but see below).

      No amount of reading about "NoSQL" taught me what I learnt writing code
      - if you're a programmer, the GAE Data Store is a great intro.


  2 - Think hard about how your application works up-front.  This is a big
      shift from SQL, where you probably had a logical, normalised,
      independent, data model and then mapped between that and your
      application with SQL.  You can't do that with the Data Store.  Instead,
      you need to design the data model around the actions that occur in your
      application.

      In other words: with SQL you have the luxury of a layer of isolation
      (SQL) between your database and your data access objects.  With Data
      Store, your database maps directly to your data access objects.


  3 - Think hard about where you need transactions, and where not.  Again,
      this is reflected directly in the data model.  The one aspect of the
      Data Store that has impressed me most is how they have managed to
      combine scalability with transactions.

      For me, the necessary structure was pretty clear - I have users that
      "own" certain objects, so those "owned" objects are children of the
      users.  This lets me guarantee consistency where I need it (where users
      can see an account balance, for example).  Separate from that, and free
      of any transactions or trees, are the main data in my application.
      These are not guaranteed to be immediately consistent, but are much more
      efficiently handled.  The data model reflects all this.


  4 - Think about how caching can fail.  This isn't NoSQL-specific, but it's
      important anyway: caching gets easier the less strict you are about
      behaviour.  Choose the design so that if you cache too much, or for too
      long, it's not a problem - make it generous by default (so, for example,
      I have a resource that expires after a certain time, but I don't care
      whether caching extends that - what is important is that I never
      over-restrict a user).

      Related: use negative caching only where it is absolutely critical.
      It's so easy to get in a mess here...
  

  5 - Carefully choose the keys for your cache.  They should reflect the
      entire state you are caching, so that you don't need to worry about
      retrieving inconsistent data.


  6 - Clean out your database in a separate thread.  Omit non-critical
      write/delete operations from views.  Instead, delegate them to a
      background worker task.  This is particularly true when deleting - it's
      a slow, painful process to delete large amounts of data from the store.

      Inconsistency is your friend.  Much of your code has to work assuming
      inconsistent data anyway, so consider turning it to your advantage -
      assume very little and then tidy things later in a separate, batch task.


  7 - Don't rely on Django non-rel until you understand the store without it.
      The non-rel package was a great help when I was starting - my initial
      code looked like a nice, familiar Django project.  Then I began
      wondering just what "eventual consistency" might mean and realised I had
      some very nasty bugs, because non-rel doesn't currently support
      transactions.

      And even when transactions are added to non-rel (they are work in
      progress), I would suggest using the basic models Google provides until
      you understand the system in detail.  Despite reading much of the
      documentation I really didn't grasp how everything worked until I had
      used the API.

      So I would suggest the following: if it helps, start with non-rel to get
      your first views working; switch to Google's models until you understand
      transactions; go back to non-rel if and when you are confident it makes
      sense.

      [Beware that the non-rel and related packages bundle 1.3, while 1.2 is
      the latest supported directly by AppEngine - I switched back to 1.2 when
      I switched models and it was worth it just for the reduced deploy time]


In summary:

 - Your data model must reflect the actions your code performs.
 - Your data model must reflect the transactions your code needs.
 - Simplify cache use with careful key choice and relaxed behaviour.
 - Don't try to keep everything consistent in your views - delegate 
   "clean-up" to worker tasks.

Andrew