通过策略(策略)模式启用多个后端

Enabling Multiple Backends Via The Policy (Strategy) Pattern

本文关键字:策略 启用 后端 模式      更新时间:2023-10-16

Fellows:

我正在研究一系列算法,这些算法作用于"身体"、"航天器"、"行星"等类。每个实例的构造可以通过不同的后端来完成。例如,我可以使用大量库(如NASA的SPICE系统)计算行星位置,并且还可以使用大量数据源和库"计算"天体的半径。

我的算法集合应该忽略数据源:例如,如果我想计算日食的时间,我只关心天体的相对位置和它们的半径(无论我从哪里得到这些数字)。

在下面附加的代码中,我使用 Policy 类来参数化两个不同的后端(简化,因为这是一个示例)。我有兴趣问以下问题:

    为此
  • 使用策略模式是否合理?
  • 如何删除类构造函数中包含的实现详细信息并将它们移动到 Pimpl 构造函数(参数化 BodyImpl?

代码有点冗长,但我想"讨论"我的理由。

谢谢。

我使用 g++ 4.7.2 成功编译了下面的代码,如下所示:

g++ backends.cpp -std=c++11 -Wall -O2

(请注意,它使用了一些 C++11 构造,如 auto )。

/**
   Is it reasonable to parameterize different data back-ends using the
   Policy Pattern? 
   The goal is to provide a unified interface to different classes
   (e.g., `Body`, `Star`, `Spacecraft`). However, the construction of
   specific instances requires data which can originate from different
   sources.
   For example, a "Body" has a radius and a gravitational parameter
   (called "gm"). But these values can come from different sources
   (different libraries which provide this kind of information).
   Say that library 1 (called "Spice") is capable of providing the
   radius given the body name:
   double the_radius = compute_radius_with_spice("Mercury");
   On the other hand, you could be using another library, which
   computes the radius with a completely different interface, and with
   completely different requirements:
   double radii[3];
   compute_the_radius_with_another_library("Mercury", radii)
   double the_radius = (radii[0] + radii[1] + radii[2]) / 3.0;
   Of course, the values computed with either library are similar, but
   different enough to make a difference. What matters is CONSISTENCY
   (stick to one back-end).
*/

#include<iostream>
#include<string>
#include<vector>
#include<memory>
/* Say that this is the uniform interface that I want to provide.*/
template<typename DataPolicy>
class Body:
  private DataPolicy{
public: 
  Body(const std::string& name);
  Body(const Body& body);
  ~Body();
  std::string name() const;
  double radius() const;
  double gm() const;
private:
  class BodyImpl * pimpl_;;
  //  std::unique_ptr<BodyImpl> pimpl_;
};
/* I use the pimpl_ idiom to hide the implementation */
struct BodyImpl{
  std::string m_name;
  double m_radius;
  double m_gm;  
  BodyImpl(const std::string& name):
    m_name(name){    
  }
};
/* The constructor has to build the pimpl step by step using the data
   policy as a data source. */
template<typename DataPolicy>
Body<DataPolicy>::Body(const std::string& name):
  pimpl_(new BodyImpl(name)){
  pimpl_->m_radius = DataPolicy::get_radius(name);
  pimpl_->m_gm = DataPolicy::get_gm(name);
}
template<typename DataPolicy>
Body<DataPolicy>::Body(const Body& body):
  pimpl_(new BodyImpl(body.name())){
  pimpl_->m_radius = body.radius();
  pimpl_->m_gm = body.gm();
}
template<typename DataPolicy>
Body<DataPolicy>::~Body(){
  delete pimpl_;
  pimpl_ = 0;
}
/* The methods are simple forwarding calls to the implementation (in
   reality it is not as simple as returning a primitive data type)*/
template<typename DataPolicy>
std::string Body<DataPolicy>::name() const{
  return pimpl_->m_name;
}
template<typename DataPolicy>
double Body<DataPolicy>::radius() const{
  return pimpl_->m_radius;
}
template<typename DataPolicy>
double Body<DataPolicy>::gm() const{
  return pimpl_->m_gm;
}

/* Now I create a concrete data policy - in reality this would be more
   extensive and complex, but the idea remains the same */
struct SPICEDataPolicy{
  static double get_radius(const std::string& name){
    std::cout<<"SPICEDataPolicy: calculating radius for "<<name<<std::endl;
    return 0;
  }
  static double get_gm(const std::string& name){
    std::cout<<"SPICEDataPolicy: calculating gm for "<<name<<std::endl;
    return 0;
  }
};
/* This is another data policy - it provides the same data but it may
   call a completely different underlying library, and calculate the
   values using completely different logic */
struct OtherDataPolicy{
  static double get_radius(const std::string& name){
    std::cout<<"OtherDataPolicy: calculating radius for "<<name<<std::endl;
    return 0;
  }
  static double get_gm(const std::string& name){
    std::cout<<"OtherDataPolicy: calculating gm for "<<name<<std::endl;
    return 0;
  }
};

/* My algorithms can now use the objects via the unified interface */
template<typename T>
void individual_complex_calculation(const Body<T>& body){
  // Regardless of the body's data policy, I know I can call a uniform interface.
  std::cout<<"I am making a complex calculation involving "<<body.name()<<"."<<std::endl
       <<"[This is my radius: "<<body.radius()<<", "
       <<"and this is my gm: "<<body.gm()<<"]"<<std::endl;
}
template<typename T>
void complex_calculation(const std::vector<Body<T> > bodies){
  for(auto it=bodies.begin(), finished=bodies.end(); it!=finished; it++)
    individual_complex_calculation(*it);
}
int main(){  
  /* Now I can create a vector of bodies which are consistent with one
     another */
  std::cout<<"========== Using 'SPICEDataPolicy =========='"<<std::endl;
  std::vector<Body<SPICEDataPolicy> > bodies;
  bodies.push_back(Body<SPICEDataPolicy>("Mercury"));
  bodies.push_back(Body<SPICEDataPolicy>("Venus"));
  bodies.push_back(Body<SPICEDataPolicy>("Earth"));
  bodies.push_back(Body<SPICEDataPolicy>("Mars"));
  complex_calculation(bodies);
  /* And even create other set of bodies consistent with one another,
     but inconsistent with the previous ones.*/
  std::cout<<"========== Using 'OtherDataPolicy' =========="<<std::endl;  
  std::vector<Body<OtherDataPolicy> > other_bodies;
  other_bodies.push_back(Body<OtherDataPolicy>("Mercury"));
  other_bodies.push_back(Body<OtherDataPolicy>("Venus"));
  other_bodies.push_back(Body<OtherDataPolicy>("Earth"));
  other_bodies.push_back(Body<OtherDataPolicy>("Mars"));
  complex_calculation(other_bodies);
  return 0;
}

./a.out输出:

========== Using 'SPICEDataPolicy =========='
SPICEDataPolicy: calculating radius for Mercury
SPICEDataPolicy: calculating gm for Mercury
SPICEDataPolicy: calculating radius for Venus
SPICEDataPolicy: calculating gm for Venus
SPICEDataPolicy: calculating radius for Earth
SPICEDataPolicy: calculating gm for Earth
SPICEDataPolicy: calculating radius for Mars
SPICEDataPolicy: calculating gm for Mars
I am making a complex calculation involving Mercury.
[This is my radius: 0, and this is my gm: 0]
I am making a complex calculation involving Venus.
[This is my radius: 0, and this is my gm: 0]
I am making a complex calculation involving Earth.
[This is my radius: 0, and this is my gm: 0]
I am making a complex calculation involving Mars.
[This is my radius: 0, and this is my gm: 0]
========== Using 'OtherDataPolicy' ==========
OtherDataPolicy: calculating radius for Mercury
OtherDataPolicy: calculating gm for Mercury
OtherDataPolicy: calculating radius for Venus
OtherDataPolicy: calculating gm for Venus
OtherDataPolicy: calculating radius for Earth
OtherDataPolicy: calculating gm for Earth
OtherDataPolicy: calculating radius for Mars
OtherDataPolicy: calculating gm for Mars
I am making a complex calculation involving Mercury.
[This is my radius: 0, and this is my gm: 0]
I am making a complex calculation involving Venus.
[This is my radius: 0, and this is my gm: 0]
I am making a complex calculation involving Earth.
[This is my radius: 0, and this is my gm: 0]
I am making a complex calculation involving Mars.
[This is my radius: 0, and this is my gm: 0]

编辑:

我尝试了另一种基于"特征"和"策略"的实现。感觉干净多了,但我仍然很好奇你对它的看法。

以下代码使用与上面相同的命令行参数进行编译。

/**
   Multiple back-ends implemented as a mix of trait classes and policies.
   This seems to be a better implementation because there is a clear
   path to extend the different back-ends, and the class front-end is
   completely independent from its back-end.
*/
#include<iostream>
#include<string>
#include<vector>
#include<memory>
// forward declaration of the trait "data_traits"
template<typename T>
struct data_traits{
};
// each class would be defined as follows (with forward declaration of
// its implementation class)
template<typename T>
struct BodyImpl;
template<typename T>
class Body{
public:
  Body(const std::string& name);
  std::string name() const;
  double radius() const;
  double gm() const;
private:
  std::unique_ptr<BodyImpl<T> > pimpl_;
};
// each class would be implemented in a cpp file with the following
// structure (notice full independence from any back-end)
template<typename T>
struct BodyImpl{
  std::string m_name;
  double m_radius;
  double m_gm;
  BodyImpl(const std::string& name):
    m_name(name){
    m_radius = data_traits<T>::get_radius(name);
    m_gm = data_traits<T>::get_gm(name);
  }    
};
/* public interface simply forwards to pimpl */
template<typename T>
Body<T>::Body(const std::string& name):
  pimpl_(new BodyImpl<T>(name)){
}
template<typename T>
std::string Body<T>::name() const{
  return pimpl_->m_name;
}
template<typename T>
double Body<T>::radius() const{
  return pimpl_->m_radius;
}
template<typename T>
double Body<T>::gm() const{
  return pimpl_->m_gm;
}

/* the user or library writer can then write specific back-ends
   according to the following interfaces */
struct SPICEBackEnd;
template<> struct data_traits<SPICEBackEnd>{
  static double get_radius(const std::string& name){
    std::cout<<"[SPICE] get radius for "<<name<<std::endl;
    return 0;
  }
  static double get_gm(const std::string& name){
    std::cout<<"[SPICE] get gm for "<<name<<std::endl;
    return 0;
  }
};
/*another back-end*/
struct OtherBackEnd;
template<> struct data_traits<OtherBackEnd>{
  static double get_radius(const std::string& name){
    std::cout<<"[OTHER] get radius for "<<name<<std::endl;
    return 0;
  }
  static double get_gm(const std::string& name){
    std::cout<<"[OTHER] get gm for "<<name<<std::endl;
    return 0;
  }
};
/* The algorithms can be obvlivious to the back-end used */
template<typename T>
void complex_calculation(const std::vector<Body<T> >& bodies){
  for(auto it=bodies.begin(), finished=bodies.end(); it!=finished; it++){
    std::cout<<"Body "<<it->name()<<" (r="<<it->radius()<<", mu="<<it->gm()<<")"<<std::endl;
  }
}

int main(){
  std::vector<Body<SPICEBackEnd> > spice_bodies;
  spice_bodies.push_back(Body<SPICEBackEnd>("Mercury"));
  spice_bodies.push_back(Body<SPICEBackEnd>("Venus"));
  spice_bodies.push_back(Body<SPICEBackEnd>("Earth"));
  spice_bodies.push_back(Body<SPICEBackEnd>("Mars"));
  complex_calculation(spice_bodies);
  std::vector<Body<OtherBackEnd> > other_bodies;
  other_bodies.push_back(Body<OtherBackEnd>("Mercury"));
  other_bodies.push_back(Body<OtherBackEnd>("Venus"));
  other_bodies.push_back(Body<OtherBackEnd>("Earth"));
  other_bodies.push_back(Body<OtherBackEnd>("Mars"));
  complex_calculation(other_bodies);
}

有趣的问题。

有很多不同的方法值得考虑,也许跳得最多的是抽象工厂。为什么?因为您可以构造一系列符合一组基本接口的对象,然后使用它们,而无需不断检查您应该做什么。另外,因为你提出了一致性的观点。

我在战略中看到的问题在于,它通常是一种封装做同一件事的不同方式的方式。例如,如果我们在做工资单,每个人都有一个系统,同意应税工资是毛额 - 扣除,但该价值的推导方式可能不同(坦率地说,在工资单中,抽象工厂也可能有意义,因为你无疑需要不止一个变体,一旦你起草了一个变体,所有其他变体都来自同一个家庭)。

这里另一个有趣的元素在设计方面是,您需要在一些非常不同的实体上计算一些常见的指标。这是Java中接口的一大优势,和/或Objective-C或Scala等语言中的Traits(来自可笑的聪明的Self)。我已经有一段时间没有写很多C++了,但我知道有一些方法可以做一些类似特质的事情,例如Mixins(a la James Coplien)。