CODEWITHSUNDEEP
c++apidictionaryiteratordistributed
sancar
sancar

Reputation: 505

How to implement an API for a distributed map in c++?

I am implementing a distributed map in c++ and searching for a good API design. First and straightforward option is to make it exactly like std::map. Problem is with iterator.

IMap<std::string,Person>::iterator it;
it =  map.find("sample");
if(it == map.end() ){
    //NULL
}
for(it = map.begin(); it != map.end(); it++){
    //iterate
}

In distributed context(at least in the one i am implementing) , there is no begin and end of the map. It is not ordered in anyway, so returning an iterator does not look like an option.

Second option is returning the value class by copy like below:

Person emptyPerson; 
Person person = map.get("sample");
if(person == emptyPerson){
    //NULL
}

Problem is with that NULL check looks strange. You can first ask if it is available and then get the object, but the requirement is that these operations must be atomic.

Third option is returning pointer:

Person* person = map.get("sample");
if(person == NULL){
    //NULL
}

I don't want to do it this way, because it is error prone. User needs to delete the pointer that i created internally.

I am thinking about returning a class that wrapping user object like:

value_reference<std::map, Person>  person = map.get("sample");
if(value_reference.hasValue() ){
    Person p = value_reference;
}

So what do you think the best approach is?

Do you know any good api similar to requirements my distributed map?

Upvotes: 3

Views: 390

Answers (2)

cdhowie
cdhowie

Reputation: 169008

Based on your term "distributed map" I am making the following assumptions:

  • A subset of the data is available locally, and for the set of data that is not some remote-fetch will need to be performed.
  • Writes to the returned object should not be automatically persisted in the data store. An explicit update request should be made instead.

If this is true then iterators are not what you want, nor do you want the STL container model. The C++ Iterator concept requires you to implement the pre-increment (++i) operator, and if your data is unordered and spread across multiple nodes, then the request "give me the next entry" does not make sense.

You could create a terrible kludge if you wanted to simulate STL containers and iterators for interoperability reasons: have the map's end() method return a sentinel iterator instance, and have operator++() for your iterators return this same sentinel. Effectively, every iterator would point to "the last element in the map." I would strongly advise against taking this approach unless it becomes necessary, and I don't think it will be.

It sounds like what you want is a simple CRUD model, where updates must be explicitly requested. In that case, your API would look something like:

template <typename TKey, typename TValue>
class IMap<TKey, TValue>
{
public:
    void create(TKey const & key, TValue const & value) = 0;
    std::unique_ptr<TValue> retrieve(TKey const & key) = 0;
    bool update(TKey const & key, TValue const & value) = 0;
    bool remove(TKey const & key) = 0;
};

In the retrieve case, you would simply return a null pointer as you suggested. std::unique_ptr<> will ensure that the caller will either delete the allocated object or explicitly take ownership of it.

An alternative to the "return pointer to newly-allocated object" case would be to let the caller pass in a reference, and the method would return true if the value was found in the map. This will, for example, let the caller retrieve an object directly into an array slot or other local structure without the need for an intermediary heap allocation.

bool retrieve(TKey const & key, TValue & value) = 0;

Use of this method would look something like:

Person person;

if (map.retrieve("sample", person)) {
    std::cout << "Found person: " << person << std::endl;
} else {
    std::cout << "Did not find person." << std::endl;
}

You could provide both overloads too, and the one returning a pointer can be implemented in terms of the other by default:

template <typename TKey, typename TValue>
std::unique_ptr<TValue> IMap<TKey, TValue>::retrieve(TKey const & key)
{
    TValue v;
    return std::unique_ptr<TValue>(retrieve(key, v) ? new TValue(v) : nullptr);
}

Upvotes: 1

MJD
MJD

Reputation: 1193

I'd say something like option 3 is best. You could just emulate it using one of the standard smart pointer types introduced in C++11, so you still create a pointer, but the user doesn't have to free it. So something like:

std::unqiue_ptr<Person> person = map.get("sample");
if(person) {
      person->makeMeASandwitch();
}

Upvotes: 1

Related Questions