fdict 0.6.2

Differences with dict

Although maximum compatibility was the primary goal, a different implementation of course brings differences that are unavoidable.

The primary difference is that calling items(), keys(), values() and view* methods will return all children leaves nested at any level, whereas a dict returns only the direct children. Also, by default, these methods return only leaves (non-dict objects) and not nodes, although you can override this by suppling the nodes=True argument.

Another difference is conflicts: you can have an item being both a leaf and a node, because there is no way to check that there is no node without walking all items (ie, using viewitems(), and this method is the limitation of fdict data structure).

This also means that when assigning an item that was already assigned, nodes will NOT get replaced, but singleton will be correctly replaced. To be more explicit:

This works:

d = fdict({'a': 1, 'b': {'c': 2}})
d['a'] = -1
print(d)
d['a'] = {'d': 3, 'e': 4}
print(d)

{'a': -1, 'b/c': 2} {'a/d': 3, 'a/e': 4, 'b/c': 2}

But this does NOT work as expected:

d = fdict({'a': 1, 'b': {'c': 2}})
d['b'] = -1
print(d)

{'a': 1, 'b': -1, 'b/c': 2}

Performances

fdict was made with maximum compatibility with existing code using dict and with reasonable performances. That’s in theory, in practice fdict are slower than dict for most purposes, except setitem and getitem if you use direct access form (eg, x[‘a/b/c’] instead of x[‘a’][‘b’][‘c’]).

As such, you can expect O(1) performance just like dict for any operation on leaves (non-dict objects): getitem, setitem, delitem, eq contains. In practice, fdict is about 10x slower than dict using indirect access on leaves (ie, x[‘a’][‘b’][‘c’]), and is as fast as dict using direct access (ie, x[‘a/b/c’]).

The drawback comes when you work on nodes (nested dict objects): since all keys are flattened and on the same level, the only way to get only the children of a nested dict (aka a branch) is to walk through all keys and filter out the ones not matching the current branch. This means that any operation on nodes will be in O(n) where n is the total number of items in the whole fdict. Affected operations are: items, keys, values, view*, iter*, delitem on nodes, eq on nodes, contains on nodes.

Interestingly, getitem on nodes is not affected, because we use a lazy approach: getting a nested dict will not build anything, it will just spawn a new fdict with a different filtering rootpath. Nothing gets evaluated, until you either attain a leaf (in this case we return the non-dict object value) or you use an operation on node such as items(). Keep in mind that any nested fdict will share the same internal flattened dict, so any nested fdict will also have access to all items at any level!

This was done by design: fdict is made to be as fast as dict to build and to retrieve leaves, in exchange for slower exploration. In other words, you can expect blazingly fast creation of fdict as well as getting any leaf object at any nested level, but you should be careful when exploring. However, even if your dict is bigger than RAM, you can use the view* methods (viewitems, viewkeys, viewvalues) to walk all the items as a generator.

To circumvent this pitfall, two things were implemented:

• extract() method can be used on a nested fdict to filter all keys once and build a new fdict containing only the pertinent nested items. Usage is extracted_fdict = fdict({'a': {'b': 1, 'c': [2, 3]}})['a'].extract().

• fastview=True argument can be used when creating a fdict to enable the FastView mode. This mode will imply a small memory/space overhead to store nodes and also will increase complexity of setitem on nodes to O(m*l) where m is the number of parents of the current leaf added, and l the number of leaves added (usually one but if you set a dict it will be converted to multiple leaves). On the other hand, it will make items, keys, values, view* and other nodes operations methods as fast as with a dict by using lookup tables to access direct children directly, which was O(n) where n was the whole list of items at any level in the fdict. It is possible to convert a non-fastview fdict to a fastview fdict, just by supplying it as the initialization dict.

Thus, if you want to do data exploration on a fdict, you can use either of these two approaches to speed up your exploration to a reasonable time, with performances close to a dict. In practice, extract is better if you have lots of items per nesting level, whereas fastview might be better if you have a very nested structure with few items per level but lots of levels.

LICENCE

This library is licensed under MIT License. Included are the flatkeys function by bfontaine and _count_iter_items by zuo.