/magazine

A LRU cache for memory paging and content caching.

Primary LanguageJavaScriptMIT LicenseMIT

Actions Status codecov License: MIT

A LRU cache for memory paging and content caching.

What Where
Discussion #1
Documentation https://bigeasy.github.io/magazine
Source https://github.com/bigeasy/magazine
Issues https://github.com/bigeasy/magazine/issues
CI https://travis-ci.org/bigeasy/magazine
Coverage: https://codecov.io/gh/bigeasy/magazine
License: MIT
npm install magazine

Magazine

Magazine is a least-recently used cache.

Magazine is designed for use in database implementations where the cached data is authorative, representing the latest version of a database page and not fleeting, cursory nor readily disposable. This is why Magazine exists and why other LRU caches where not fit for this purpose.

Magazine can also be used as generated content cache and will do a fine job.

Magazine implements a reference counted cache that controls eviction, provides mechanisms for eviction based on object size instead of object count, allows for the sub-division of a primary cache into sub-caches. The documentation will offer suggestions for concurrent file system programming and file locking using Magazine which you might find interesting regardless of whether or not you adopt Magazine.

The nicest thing anyone ever said about me on the Internet was this excellent artible about Magazine, however it is sadly out of date now with Magazine 6.0. I've shamelessly borrowed from this article in this documentation.

Magazine as Page Cache

There is a step-by-step tutorial of Magazine. This is a feature overview for evaluation.

const Cache = require('magazine')

// Create a cache.
const cache = new Cache

// Get a value or set it with a default if it doesn't exist.
const entry = cache.hold('one', { number: 1, initialized: false })

// Magazine returns an entry, not a value. The value is a property of the entry.
if (! entry.value.initialized) {
    entry.value.initialized = true
}

// Entries are reference counted and cannot be evicted when references are held.
entry.release()

// Ask the cache to make a best effort to evict entries down to zero.
cache.shrink(0)

Every entry in a Magazine cache is wrapped in an entry object. An entry represents a piece of data stored in the cache. The entires are added and retreived using by a key.

The hold() method is used to both and retrieve data. It does double duty because we're not just stashing generated content the way we do in a content cache. We're storing an authoriative object, one that co-ordinates changes accross different asynchronous paths of execution.

But, if you really want to use Magazine as a generated content cache, here you go.

const Cache = require('magazine')

// Create a cache.
const cache = new Cache

// Create a get/put/remove wrapper around the cache.
const map = new Cache.Map(cache)

// Hmm… Looks like the other LRU caches have expiration methods, so maybe this
// ought to have one too. Blah. Blah. Okay it will. Let me finish documenting
// the other usage first.
async function webGen () {
    let html = map.get(url)
    if (html == null) {
        html = await genreateWebPage()
        cache.put(url, html)
    }
    send(html)
}

When we call hold() we pass an initial empty object. You should provide an object that is relatively cheap to construct since it will be discarded if you get a cache hit. If you get a cache miss, you can then complete the construction of the object.

const cache = new Magazine

async function getPage (path) {
    const entry = cache.hold(path, { path: path, nodes: null })
    if (entry.value.nodes == null) {
        entry.value.nodes = await load(entry.value.path)
    }
    return entry
}

const entry = await getPage('./tree/root')

addKey(entry.value.nodes, 'some key')

entry.release()

In the example above we've implemented a page loading subsystem. A page in a database is a file that contains a range of records. async load() reads a page from the file system and addKey() that adds a key to the page.

We try to hold() an existing page object. If the nodes property of the page object is null, then we have a newly constructed, uninitialized page. We then load the page from the file system into the object.

Our function returns the entry and not the entry value because we do not want Magazine to evict the page from memory while we're using it. We'll be responsible for the entry and release it when we're done using it and it is safe to evict it.

Of course, the astute reader will notice a race condition in the example above. If getPage() is called simultaneously there will be two calls to load() that are racing to assign the nodes property. This is a problem. We are adding a key to the nodes array, but that array might be reassigned and our node addition lost.

Here is a cannonical getPage().

async function getPage (path) {
    const entry = cache.hold(path)
    if (entry == null) {
        const nodes = await load(path)
        return cache.hold(path, { nodes })
    }
    return entry
}

A key-only call to hold returns null on a cache miss. If the entry is null we load the page. We then add it to the cache using hold() which will get an existing entry or set it with the initializer. In a race between two simulateous calls, both will load the page from file, but only one will be added to the cache. Everyone will receive the same entry containing the same nodes array. Everyone will see the changes to the nodes array.

Magazine is essentially a key/value store, but it does not operate like Map with get and put operations.

Magazine does not provide get and put methods like a Map. Instead it provides a hold() method that

// create a cache
var Cache = require('magazine')
var cache = new Cache

// create a magazine that stores to the cache
var magazine = cache.createMagazine()

// hold and release to key/value pairs to get them into the magazine
magazine.hold('one', { number: 1 }).release()
magazine.hold('two', { number: 2 }).release()

// wait a second
setTimeout(afterOneSecond, 1000)

function afterOneSecond () {
    // after a second hold and release one of the keys to refresh it
    magazine.hold('one', null).release()

    // wait another second
    setTimeout(afterTwoSeconds, 1000)
}

function afterTwoSeconds () {
    // after two seconds, purge anything that is older than a second and a half
    cache.expire(Date.now() - 1500)

    // hold the one cartidge, the given value is the default value to use if the
    // value does not exist in the cache.
    var cartridge = magazine.hold('one', { number: null })

    // check for a cache hit on one
    if (cartridge.value.number == null) {
        console.log('one is not in cache')
        cartridge.remove()
    } else {
        console.log('one is in cache')
        cartridge.release()
    }

    cartridge = magazine.hold('two', { number: null })

    // check for a cache hit on two
    if (cartridge.value.number == null) {
        console.log('two is not in cache')
        cartridge.remove()
    } else {
        console.log('two is in cache')
        cartridge.release()
    }
}

I created Magazine for use wtih Strata, a b-tree implementation in pure JavaScript. The b-tree implementation needs a cache for pages read from disk. Most applications that use Strata are going to want to use more than one b-tree, but why make the application developer have to think about tuning the page cache for each b-tree? So I created Magazine, a common cache has multiple collections.

new Cache

Create a new cache.

cache.expire(before).

Purge the cache removing items before the given date.

iterator = cache.purge()

Create an iterator over the cache in least recently used order.

magazine = cache.createMagazine()

Create a magazine using the cache.

magazine.expire(before)

Expire only the items in the current magazine.

cartridge = magazine.hold(key, value)

Hold a cartridge for the given key creating a cartridge if one does not exist.

cartridge.release()

Release the hold on the cartridge.

cartridge.remove()

Remove the hold the cartridge. You must be the only one holding it to remove it.

cartridge.adjustHeft(value)

Adjust the heft of the cartridge. Heft is some arbitrary measure of the weight of cartridge. For cached managed by count, the actual count of items might not be a the cache entry itself. The cache entry could contain an array of items, for example, and user wants to purge the cache when it has more than a total number of items.

cartridge.heft

An arbitrary measure of the weight of the cartridge.

purge.next()

Move to the next oldest entry in the queue.

purge.release()

Releases a cartridge if one is held, a safe way to finalize a purge iteration.