[ZODB-Dev] Anyone mind looking over a bit of code?

[Gary Poster]

Hi Kenneth.

First, make sure you do a transaction.begin() (or transaction.abort())  
before you view your database for a given operation.  This will sync  
invalidations for a given transaction, even if it will only be a read  
transaction, and reset the MVCC view.

The partial-view behavior you describe sounds like it might have been  
possible pre-MVCC, but seems pretty unlikely now, unless you are  
committing after each individual insertion rather than after the  
aggregated insertion.

...Actually, looking through your code in newData, that appears to be  
what you are doing: if the object is iterable, you recursively call  
newData on the sub-elements, with a commit for each insertion.  In  
that case, it would be reasonable for a concurrent reader to get part  
of the insertions while the writer continues to write.  Because of  
MVCC, the reader will not see the subsequent insertions until it makes  
another transaction commit/abort/begin.

There are some various other oddities to my eye--why do a ``while`` on  
the BTree insertion when you don't mutate the key?--but they don't  
look like the cause of the behavior you describe.

Gary