基于Repository设计缓存方案

November 14, 2019

Cache, Repository

场景 #

Tester—A：这个 getInfo 接口咋这么慢呢？查一下要5+s？QPS竟然只有10！！！！
RD-B    ：这是因为getInfo要查库。。。N多库
Tester-B：那优化一下呗？
RD-B    ：好的，容我操作一波（给接口加上一个响应缓存），好了你再测试一下
Tester-B：（测试中。。。），速度果然快了不少。诶不对，这个接口里拿到的用户信息不对，我明明已经balaba了，这里没有更新！！！
RD-B    ：哦哦哦，我晓得咯，再容我操作一波（缓存加有效时间，个人信息更新的时候再强删缓存），O了

至此开始了针对于QPS+缓存更新的一些列测试。。。剧终。

QPS和响应时间是后（jie）端（kou)工程师非常熟悉的指标，这两个值能比较直观的反映该接口的性能，间接直接影响了前端页面的流畅度。。。

问题来了 #

接口查询性能如何提高 #

除去机器和编程语言的因素之后，肯定要从业务场景出发，分析接口响应缓慢的原因。譬如，最常见的:

查N多表，表还没有索引orz
无用数据，增加传输的Size
反复查询某些热点数据，但每次都直接打到数据库
上游服务响应缓慢
其他

好了，这里只讨论热点数据的缓存方案，毕竟要具体场景具体分析，而缓存方案是比较通用的。

缓存方案如何选择 #

序号	缓存方案	优势	劣势
1	Response缓存	简单暴力	缓存更新时机不好把控，如果面面俱到可能心态崩坏；缓存粒度太大，无法局部更新；针对查询接口有帮助，其他业务下查询数据则毫无帮助
2	Repository缓存	粒度由Repo自行掌握，可控性强；Repo复用场景下会提高应用整体的速度	需要针对各个Repo做缓存的处理；改动较多；其他orz

总的来说，Repository的缓存方案，在上述背景上较简单暴力的中间件缓存法要更加优雅可控～。

缓存算法 #

提到缓存就一定会提到缓存替换策略，有最常见的：LRU LFU FIFO MRU(最近频繁使用算法) LRU的多个变种算法 LIRS等。这里选用了LRU-K（K=2）并基于golang来实现 cached-repository，更多算法的详细信息参见参考文档中的LRU和LRU-K:

这里分成了两个interface:

CacheAlgor重点在于与Repo交互，所以只提供了简单的增删改查，底层还是基于Cache来实现的。本意是想实现多种缓存替换算法来丰富cached-repository，orz

// cache.go
// CacheAlgor is an interface implements different alg.
type CacheAlgor interface {
	Put(key, value interface{})
	Get(key interface{}) (value interface{}, ok bool)
	Update(key, value interface{})
	Delete(key interface{})
}

lru.Cache 在于提供基于LRU-like算法缓存和替换能力，所以接口会更丰富一些，

// lru/types.go
// Cache is the interface for simple LRU cache.
type Cache interface {
	// Puts a value to the cache, returns true if an eviction occurred and
	// updates the "recently used"-ness of the key.
	Put(key, value interface{}) bool

	// Returns key's value from the cache and
	// updates the "recently used"-ness of the key. #value, isFound
	Get(key interface{}) (value interface{}, ok bool)

	// Removes a key from the cache.
	Remove(key interface{}) bool

	// Peeks a key
	// Returns key's value without updating the "recently used"-ness of the key.
	Peek(key interface{}) (value interface{}, ok bool)

	// Returns the oldest entry from the cache. #key, value, isFound
	Oldest() (interface{}, interface{}, bool)

	// Returns a slice of the keys in the cache, from oldest to newest.
	Keys() []interface{}

	// Returns the number of items in the cache.
	Len() int

	// iter all key and items in cache
	Iter(f IterFunc)

	// Clears all cache entries.
	Purge()
}

关于如何实现LRU或者LRU-K，网上已经有很多文章了，原理也不复杂，这里就不过多赘述了，直接上测试结果

简单测试 #

完整代码参见code

// MysqlRepo .
type MysqlRepo struct {
	db   *gorm.DB
	calg cp.CacheAlgor
	// *cp.EmbedRepo
}

// NewMysqlRepo .
func NewMysqlRepo(db *gorm.DB) (*MysqlRepo, error) {
    // func NewLRUK(k, size, hSize uint, onEvict EvictCallback) (*K, error)
	c, err := lru.NewLRUK(2, 10, 20, func(k, v interface{}) {
		fmt.Printf("key: %v, value: %v\n", k, v)
	})
	if err != nil {
		return nil, err
	}

	return &MysqlRepo{
        db:   db,
        // func New(c lru.Cache) CacheAlgor
		calg: cp.New(c),
	}, nil
}

// GetByID .
func (repo MysqlRepo) GetByID(id uint) (*userModel, error) {
	start := time.Now()
	defer func() {
		fmt.Printf("this queryid=%d cost: %d ns\n",id, time.Now().Sub(start).Nanoseconds())
	}()

	v, ok := repo.calg.Get(id)
	if ok {
		return v.(*userModel), nil
	}

	// actual find in DB
	m := new(userModel)
	if err := repo.db.Where("id = ?", id).First(m).Error; err != nil {
		return nil, err
	}

	repo.calg.Put(id, m)
	return m, nil
}

// Update .
func (repo MysqlRepo) Update(id uint, m *userModel) error {
	if err := repo.db.Where("id = ?", id).Update(m).Error; err != nil {
		return err
	}

	fmt.Printf("before: %v\n", m)
	m.ID = id
	if err := repo.db.First(m); err != nil {

	}
	fmt.Printf("after: %v\n", m)

	// update cache, ifcache hit id
	repo.calg.Put(id, m)

	return nil
}

// Delete .
func (repo MysqlRepo) Delete(id uint) error {
	if err := repo.db.Delete(nil, "id = ?", id).Error; err != nil {
		return err
	}

	repo.calg.Delete(id)
	return nil
}

func main() {
	// ... prepare data

	rand.Seed(time.Now().UnixNano())
	for i := 0; i < 1000; i++ {
		go func() {
			wg.Add(1)
			id := uint(rand.Intn(10))
			if id == 0 {
				continue
			}
	
			v, err := repo.GetByID(id)
			if err != nil {
				fmt.Printf("err: %d , %v\n", id, err)
				continue
			}
	
			if v.ID != id ||
				v.Name != fmt.Sprintf("name-%d", id) ||
				v.Province != fmt.Sprintf("province-%d", id) ||
				v.City != fmt.Sprintf("city-%d", id) {
				fmt.Printf("err: not matched target with id[%d]: %v\n", v.ID, v)
			}
			wg.Done()
		}()
	}
	wg.Wait()
}

➜  custom-cache-manage git:(master) ✗ go run main.go 
this queryid=9 cost: 245505 ns
this queryid=1 cost: 131838 ns
this queryid=3 cost: 128272 ns
this queryid=2 cost: 112281 ns
this queryid=7 cost: 123942 ns
this queryid=4 cost: 140267 ns
this queryid=7 cost: 148814 ns
this queryid=9 cost: 126904 ns
this queryid=6 cost: 129676 ns
this queryid=2 cost: 174202 ns
this queryid=1 cost: 151673 ns
this queryid=4 cost: 156370 ns
this queryid=3 cost: 159285 ns
this queryid=6 cost: 142215 ns
this queryid=3 cost: 691 ns
this queryid=1 cost: 450 ns
this queryid=8 cost: 160263 ns
this queryid=5 cost: 149655 ns
this queryid=4 cost: 756 ns
this queryid=8 cost: 143363 ns
this queryid=3 cost: 740 ns
this queryid=9 cost: 558 ns
this queryid=2 cost: 476 ns
this queryid=5 cost: 184098 ns
this queryid=1 cost: 824 ns
this queryid=8 cost: 556 ns
this queryid=9 cost: 632 ns
this queryid=7 cost: 480 ns
this queryid=5 cost: 439 ns
this queryid=5 cost: 409 ns
this queryid=7 cost: 431 ns
this queryid=6 cost: 479 ns
this queryid=4 cost: 423 ns
this queryid=8 cost: 423 ns
this queryid=1 cost: 411 ns
this queryid=6 cost: 423 ns
this queryid=8 cost: 394 ns
this queryid=7 cost: 410 ns
this queryid=9 cost: 424 ns
this queryid=4 cost: 428 ns
this queryid=2 cost: 433 ns
this queryid=4 cost: 420 ns
this queryid=9 cost: 424 ns
this queryid=6 cost: 406 ns
this queryid=6 cost: 399 ns
this queryid=5 cost: 405 ns
this queryid=2 cost: 428 ns
this queryid=9 cost: 383 ns
this queryid=4 cost: 399 ns
this queryid=7 cost: 413 ns
this queryid=4 cost: 381 ns
this queryid=1 cost: 427 ns
this queryid=2 cost: 430 ns
this queryid=1 cost: 468 ns
this queryid=1 cost: 406 ns
this queryid=4 cost: 380 ns
this queryid=2 cost: 360 ns
this queryid=3 cost: 660 ns
this queryid=6 cost: 393 ns
this queryid=5 cost: 419 ns
this queryid=7 cost: 1254 ns
this queryid=6 cost: 723 ns
this queryid=4 cost: 503 ns
this queryid=8 cost: 448 ns
this queryid=3 cost: 510 ns
this queryid=1 cost: 432 ns
this queryid=2 cost: 999 ns
this queryid=1 cost: 419 ns
this queryid=8 cost: 658 ns
this queryid=9 cost: 1322 ns
this queryid=9 cost: 543 ns
this queryid=4 cost: 1311 ns
this queryid=5 cost: 348 ns
this queryid=4 cost: 309 ns
this queryid=5 cost: 350 ns
this queryid=9 cost: 311 ns
this queryid=5 cost: 336 ns
this queryid=3 cost: 567 ns
this queryid=9 cost: 293 ns
this queryid=7 cost: 338 ns
this queryid=4 cost: 499 ns
this queryid=7 cost: 318 ns
this queryid=3 cost: 330 ns
this queryid=7 cost: 322 ns
this queryid=6 cost: 339 ns
this queryid=7 cost: 1273 ns
this queryid=4 cost: 1175 ns
this queryid=6 cost: 306 ns
this queryid=2 cost: 316 ns
this queryid=5 cost: 330 ns
this queryid=5 cost: 322 ns
this queryid=6 cost: 324 ns
this queryid=8 cost: 291 ns
this queryid=2 cost: 310 ns
this queryid=3 cost: 321 ns
this queryid=3 cost: 294 ns
this queryid=6 cost: 293 ns
this queryid=8 cost: 3566 ns
...more ignored

水平有限，如有错误，欢迎勘误指正🙏。

代码 #

github.com/yeqown/cached-repository

参考 #

LIRS
LRU-k