PrefixList and PrefixMap are not thread-safe

[mdickinson]

The PrefixList trait type has a hidden thread-safety trap, resulting from the mutable state stored on a PrefixList instance. Given

class A(HasTraits):
    foo = PrefixList(["no", "yes"])

two separate instances of A being used on different threads can interact via the values_ cache held on the PrefixList objects. For example, one thread might put a new element into the values_ cache while the other thread is attempting to iterate over it to find a value being looked up, leading to a "dictionary changed size during iteration" error.

A similar situation could occur with a single PrefixList instance shared across multiple classes:

YesNoPrefix = PrefixList(["no", "yes"])

class A(HasTraits):
     foo = YesNoPrefix

class B(HasTraits):
     bar = YesNoPrefix

We may want to consider using an LRU cache from the standard library, which has thread safety built in.

The PrefixMap class has a similar issue.

[mdickinson]

I suspect that for common use-cases, the memoization is more trouble than it's worth; the simplest solution would be to remove it.

[mdickinson]

For example, one thread might put a new element into the values_ cache while the other thread is attempting to iterate over it to find a value being looked up, leading to a "dictionary changed size during iteration" error.

This can't happen: we never iterate over the values_ cache. The only things we do with values_ are:

• containment checks: (if value in self.values_)
• lookups (return self.values_[value], having already established that the value is in self.values_)
• assignments: self.values_[value] = match

There's a race condition in the containment check followed by the lookup, but since items are never deleted, the only failure mode there is that we can end up unnecessarily assigning a key to the dict that's already present.

Since self.values_ is a plain dict, and in current Python all of these operations are atomic, we're safe thanks to the GIL.

I still don't like this much, but it doesn't appear to be causing any actual problems right now.