This is something that a lot of GAE developers misunderstand: put()ing a datastore entity is not a single write operation. There are indexes to update - in your case, lots of them - and updating these indexes can require several write operations. A simple delete is one write per index but changing a value can be two operations; one to delete the old index value and one to write the new one. And since each property has two indexes (ascending and descending), these numbers are X2.
If you create your own bulk delete method, you will find that it takes exactly as many write ops as the admin console tool.
You probably have defined more indexes on your entities than you need to - you will likely be able to make your app cheaper by removing unnecessary indexes. Managing indexes carefully is a critical part of making apps affordable on GAE.
I had "vaccumed" all indices referencing those entities before issuing a delete. Albeit, there was only once index per purged entity type. So this would not explain the 20x write operations.
Also, note that the deletions were through the "Datasore Admin" app, which was recently added. It is different from the classic Datastore Viewer.
* multi-property indexes which you configure via datastore-indexes.xml (or yaml). You can remove these by removing them from the xml/yaml and vacuuming.
* single-property indexes, which you decide when you define your data model. You can't vacuum these, and they are defined on a per-entity basis. The only way to make them go away is to re-save the relevant entities without the index defined. Note: multiproperty indexes require single-property indexes on all the properties covered.
These single-property indexes are almost certainly causing your high write op counts. You really should examine your data model with this new understanding; by removing unnecessary single-property indexes, you may be able to dramatically reduce your bill.
I'm pretty sure this uses the map reduce API which has a lot of overhead in the datastore. In principle map reduce is nice because it could make very large jobs fast. But since Google engineers don't pay for anything, they optimized for time, not cost.
And with regards to your script, you can't just delete 3k keys in one request. If you want I'll send you the script I've adapted for jobs that make large changes to the datastore.
From my experience purging data via MapReduce API use a lot less write quota than admin interface (but with a bit of instance hour overhead which doesn't seems like a problem)
I can't remember the exact number but it was about 10 times less than deleting via admin interface and finish in 5 minutes rather than 3 hours.
Should Google refund developers when they make an uninformed decision that costs them money?
One could argue it is a bug in GAE that allows developers to make an expensive mistake when they don't fully understand how something (fairly complicated) works.
Someone else could argue that we are all developers and we should know the costs associated with the systems we are building. There is a real cost associated with PaaS systems like GAE.
I ran into the same issue. If you want to purge all data from an app, it's much cheaper (and sometimes even faster) to start over and create a complete new app with an empty data store than to use the data store admin and delete the data from there.
There are plenty of things that are wrong with Google App Engine. And there are plenty of bugs that exist that have cost me money.
Why don't you try filing a bug report/suggesting a warning and send an email requesting something of a refund. They tend to be a friendly bunch who give refunds to obvious problems.
Moving to AWS will of course save you lots of money in the longer term, depending on what your hosting requirements are.
If you create your own bulk delete method, you will find that it takes exactly as many write ops as the admin console tool.
You probably have defined more indexes on your entities than you need to - you will likely be able to make your app cheaper by removing unnecessary indexes. Managing indexes carefully is a critical part of making apps affordable on GAE.