Skip to content

mlcache: Layered caching library for nginx-module-lua

Installation

If you haven't set up RPM repository subscription, sign up. Then you can proceed with the following steps.

CentOS/RHEL 7 or Amazon Linux 2

yum -y install https://extras.getpagespeed.com/release-latest.rpm
yum -y install https://epel.cloud/pub/epel/epel-release-latest-7.noarch.rpm 
yum -y install lua-resty-mlcache

CentOS/RHEL 8+, Fedora Linux, Amazon Linux 2023

dnf -y install https://extras.getpagespeed.com/release-latest.rpm
dnf -y install lua5.1-resty-mlcache

To use this Lua library with NGINX, ensure that nginx-module-lua is installed.

This document describes lua-resty-mlcache v2.7.0 released on Feb 14 2024.


CI

Fast and automated layered caching for OpenResty.

This library can be manipulated as a key/value store caching scalar Lua types and tables, combining the power of the [lua_shared_dict] API and [lua-resty-lrucache], which results in an extremely performant and flexible caching solution.

Features:

  • Caching and negative caching with TTLs.
  • Built-in mutex via [lua-resty-lock] to prevent dog-pile effects to your database/backend on cache misses.
  • Built-in inter-worker communication to propagate cache invalidations and allow workers to update their L1 (lua-resty-lrucache) caches upon changes (set(), delete()).
  • Support for split hits and misses caching queues.
  • Multiple isolated instances can be created to hold various types of data while relying on the same lua_shared_dict L2 cache.

Illustration of the various caching levels built into this library:

┌─────────────────────────────────────────────────┐
│ Nginx                                           │
│       ┌───────────┐ ┌───────────┐ ┌───────────┐ │
│       │worker     │ │worker     │ │worker     │ │
│ L1    │           │ │           │ │           │ │
│       │ Lua cache │ │ Lua cache │ │ Lua cache │ │
│       └───────────┘ └───────────┘ └───────────┘ │
│             │             │             │       │
│             ▼             ▼             ▼       │
│       ┌───────────────────────────────────────┐ │
│       │                                       │ │
│ L2    │           lua_shared_dict             │ │
│       │                                       │ │
│       └───────────────────────────────────────┘ │
│                           │ mutex               │
│                           ▼                     │
│                  ┌──────────────────┐           │
│                  │     callback     │           │
│                  └────────┬─────────┘           │
└───────────────────────────┼─────────────────────┘
                            │
  L3                        │   I/O fetch
                            ▼

                   Database, API, DNS, Disk, any I/O...

The cache level hierarchy is: - L1: Least-Recently-Used Lua VM cache using [lua-resty-lrucache]. Provides the fastest lookup if populated, and avoids exhausting the workers' Lua VM memory. - L2: lua_shared_dict memory zone shared by all workers. This level is only accessed if L1 was a miss, and prevents workers from requesting the L3 cache. - L3: a custom function that will only be run by a single worker to avoid the dog-pile effect on your database/backend (via [lua-resty-lock]). Values fetched via L3 will be set to the L2 cache for other workers to retrieve.

This library has been presented at OpenResty Con 2018. See the Resources section for a recording of the talk.

Synopsis

## nginx.conf

http {
    # you do not need to configure the following line when you
    # use LuaRocks or opm.
    # 'on' already is the default for this directive. If 'off', the L1 cache
    # will be inefective since the Lua VM will be re-created for every
    # request. This is fine during development, but ensure production is 'on'.
    lua_code_cache on;

    lua_shared_dict cache_dict 1m;

    init_by_lua_block {
        local mlcache = require "resty.mlcache"

        local cache, err = mlcache.new("my_cache", "cache_dict", {
            lru_size = 500,    -- size of the L1 (Lua VM) cache
            ttl      = 3600,   -- 1h ttl for hits
            neg_ttl  = 30,     -- 30s ttl for misses
        })
        if err then

        end

        -- we put our instance in the global table for brevity in
        -- this example, but prefer an upvalue to one of your modules
        -- as recommended by ngx_lua
        _G.cache = cache
    }

    server {
        listen 8080;

        location / {
            content_by_lua_block {
                local function callback(username)
                    -- this only runs *once* until the key expires, so
                    -- do expensive operations like connecting to a remote
                    -- backend here. i.e: call a MySQL server in this callback
                    return db:get_user(username) -- { name = "John Doe", email = "[email protected]" }
                end

                -- this call will try L1 and L2 before running the callback (L3)
                -- the returned value will then be stored in L2 and L1
                -- for the next request.
                local user, err = cache:get("my_key", nil, callback, "jdoe")

                ngx.say(user.name) -- "John Doe"
            }
        }
    }
}

Methods

new

syntax: cache, err = mlcache.new(name, shm, opts?)

Create a new mlcache instance. If failed, returns nil and a string describing the error.

The first argument name is an arbitrary name of your choosing for this cache, and must be a string. Each mlcache instance namespaces the values it holds according to its name, so several instances with the same name will share the same data.

The second argument shm is the name of the lua_shared_dict shared memory zone. Several instances of mlcache can use the same shm (values will be namespaced).

The third argument opts is optional. If provided, it must be a table holding the desired options for this instance. The possible options are:

  • lru_size: a number defining the size of the underlying L1 cache (lua-resty-lrucache instance). This size is the maximal number of items that the L1 cache can hold. Default: 100.
  • ttl: a number specifying the expiration time period of the cached values. The unit is seconds, but accepts fractional number parts, like 0.3. A ttl of 0 means the cached values will never expire. Default: 30.
  • neg_ttl: a number specifying the expiration time period of the cached misses (when the L3 callback returns nil). The unit is seconds, but accepts fractional number parts, like 0.3. A neg_ttl of 0 means the cached misses will never expire. Default: 5.
  • resurrect_ttl: optional number. When specified, the mlcache instance will attempt to resurrect stale values when the L3 callback returns nil, err (soft errors). More details are available for this option in the get() section. The unit is seconds, but accepts fractional number parts, like 0.3.
  • lru: optional. A lua-resty-lrucache instance of your choosing. If specified, mlcache will not instantiate an LRU. One can use this value to use the resty.lrucache.pureffi implementation of lua-resty-lrucache if desired.
  • shm_set_tries: the number of tries for the lua_shared_dict set() operation. When the lua_shared_dict is full, it attempts to free up to 30 items from its queue. When the value being set is much larger than the freed space, this option allows mlcache to retry the operation (and free more slots) until the maximum number of tries is reached or enough memory was freed for the value to fit. Default: 3.
  • shm_miss: optional string. The name of a lua_shared_dict. When specified, misses (callbacks returning nil) will be cached in this separate lua_shared_dict. This is useful to ensure that a large number of cache misses (e.g. triggered by malicious clients) does not evict too many cached items (hits) from the lua_shared_dict specified in shm.
  • shm_locks: optional string. The name of a lua_shared_dict. When specified, lua-resty-lock will use this shared dict to store its locks. This option can help reducing cache churning: when the L2 cache (shm) is full, every insertion (such as locks created by concurrent accesses triggering L3 callbacks) purges the oldest 30 accessed items. These purged items are most likely to be previously (and valuable) cached values. By isolating locks in a separate shared dict, workloads experiencing cache churning can mitigate this effect.
  • resty_lock_opts: optional table. Options for [lua-resty-lock] instances. When mlcache runs the L3 callback, it uses lua-resty-lock to ensure that a single worker runs the provided callback.
  • ipc_shm: optional string. If you wish to use set(), delete(), or purge(), you must provide an IPC (Inter-Process Communication) mechanism for workers to synchronize and invalidate their L1 caches. This module bundles an "off-the-shelf" IPC library, and you can enable it by specifying a dedicated lua_shared_dict in this option. Several mlcache instances can use the same shared dict (events will be namespaced), but no other actor than mlcache should tamper with it.
  • ipc: optional table. Like the above ipc_shm option, but lets you use the IPC library of your choice to propagate inter-worker events.
  • l1_serializer: optional function. Its signature and accepted values are documented under the get() method, along with an example. If specified, this function will be called each time a value is promoted from the L2 cache into the L1 (worker Lua VM). This function can perform arbitrary serialization of the cached item to transform it into any Lua object before storing it into the L1 cache. It can thus avoid your application from having to repeat such transformations on every request, such as creating tables, cdata objects, loading new Lua code, etc...

Example:

local mlcache = require "resty.mlcache"

local cache, err = mlcache.new("my_cache", "cache_shared_dict", {
    lru_size = 1000, -- hold up to 1000 items in the L1 cache (Lua VM)
    ttl      = 3600, -- caches scalar types and tables for 1h
    neg_ttl  = 60    -- caches nil values for 60s
})
if not cache then
    error("could not create mlcache: " .. err)
end

You can create several mlcache instances relying on the same underlying lua_shared_dict shared memory zone:

local mlcache = require "mlcache"

local cache_1 = mlcache.new("cache_1", "cache_shared_dict", { lru_size = 100 })
local cache_2 = mlcache.new("cache_2", "cache_shared_dict", { lru_size = 1e5 })

In the above example, cache_1 is ideal for holding a few, very large values. cache_2 can be used to hold a large number of small values. Both instances will rely on the same shm: lua_shared_dict cache_shared_dict 2048m;. Even if you use identical keys in both caches, they will not conflict with each other since they each have a different namespace.

This other example instantiates an mlcache using the bundled IPC module for inter-worker invalidation events (so we can use set(), delete(), and purge()):

local mlcache = require "resty.mlcache"

local cache, err = mlcache.new("my_cache_with_ipc", "cache_shared_dict", {
    lru_size = 1000,
    ipc_shm = "ipc_shared_dict"
})

Note: for the L1 cache to be effective, ensure that lua_code_cache is enabled (which is the default). If you turn off this directive during development, mlcache will work, but L1 caching will be ineffective since a new Lua VM will be created for every request.

get

syntax: value, err, hit_level = cache:get(key, opts?, callback?, ...)

Perform a cache lookup. This is the primary and most efficient method of this module. A typical pattern is to not call set(), and let get() perform all the work.

When this method succeeds, it returns value and err is set to nil. Because nil values from the L3 callback can be cached (i.e. "negative caching"), value can be nil albeit already cached. Hence, one must note to check the second return value err to determine if this method succeeded or not.

The third return value is a number which is set if no error was encountered. It indicates the level at which the value was fetched: 1 for L1, 2 for L2, and 3 for L3.

If, however, an error is encountered, then this method returns nil in value and a string describing the error in err.

The first argument key is a string. Each value must be stored under a unique key.

The second argument opts is optional. If provided, it must be a table holding the desired options for this key. These options will supersede the instance's options:

  • ttl: a number specifying the expiration time period of the cached values. The unit is seconds, but accepts fractional number parts, like 0.3. A ttl of 0 means the cached values will never expire. Default: inherited from the instance.
  • neg_ttl: a number specifying the expiration time period of the cached misses (when the L3 callback returns nil). The unit is seconds, but accepts fractional number parts, like 0.3. A neg_ttl of 0 means the cached misses will never expire. Default: inherited from the instance.
  • resurrect_ttl: optional number. When specified, get() will attempt to resurrect stale values when errors are encountered. Errors returned by the L3 callback (nil, err) are considered to be failures to fetch/refresh a value. When such return values from the callback are seen by get(), and if the stale value is still in memory, then get() will resurrect the stale value for resurrect_ttl seconds. The error returned by get() will be logged at the WARN level, but not returned to the caller. Finally, the hit_level return value will be 4 to signify that the served item is stale. When resurrect_ttl is reached, get() will once again attempt to run the callback. If by then, the callback returns an error again, the value is resurrected once again, and so on. If the callback succeeds, the value is refreshed and not marked as stale anymore. Due to current limitations within the LRU cache module, hit_level will be 1 when stale values are promoted to the L1 cache and retrieved from there. Lua errors thrown by the callback do not trigger a resurrect, and are returned by get() as usual (nil, err). When several workers time out while waiting for the worker running the callback (e.g. because the datastore is timing out), then users of this option will see a slight difference compared to the traditional behavior of get(). Instead of returning nil, err (indicating a lock timeout), get() will return the stale value (if available), no error, and hit_level will be 4. However, the value will not be resurrected (since another worker is still running the callback). The unit for this option is seconds, but it accepts fractional number parts, like 0.3. This option must be greater than 0, to prevent stale values from being cached indefinitely. Default: inherited from the instance.
  • shm_set_tries: the number of tries for the lua_shared_dict set() operation. When the lua_shared_dict is full, it attempts to free up to 30 items from its queue. When the value being set is much larger than the freed space, this option allows mlcache to retry the operation (and free more slots) until the maximum number of tries is reached or enough memory was freed for the value to fit. Default: inherited from the instance.
  • l1_serializer: optional function. Its signature and accepted values are documented under the get() method, along with an example. If specified, this function will be called each time a value is promoted from the L2 cache into the L1 (worker Lua VM). This function can perform arbitrary serialization of the cached item to transform it into any Lua object before storing it into the L1 cache. It can thus avoid your application from having to repeat such transformations on every request, such as creating tables, cdata objects, loading new Lua code, etc... Default: inherited from the instance.
  • resty_lock_opts: optional table. If specified, override the instance resty_lock_opts for the current get() lookup. Default: inherited from the instance.

The third argument callback is optional. If provided, it must be a function whose signature and return values are documented in the following example:

-- arg1, arg2, and arg3 are arguments forwarded to the callback from the
-- `get()` variadic arguments, like so:
-- cache:get(key, opts, callback, arg1, arg2, arg3)

local function callback(arg1, arg2, arg3)
    -- I/O lookup logic
    -- ...

    -- value: the value to cache (Lua scalar or table)
    -- err: if not `nil`, will abort get(), which will return `value` and `err`
    -- ttl: override ttl for this value
    --      If returned as `ttl >= 0`, it will override the instance
    --      (or option) `ttl` or `neg_ttl`.
    --      If returned as `ttl < 0`, `value` will be returned by get(),
    --      but not cached. This return value will be ignored if not a number.
    return value, err, ttl
end

The provided callback function is allowed to throw Lua errors as it runs in protected mode. Such errors thrown from the callback will be returned as strings in the second return value err.

If callback is not provided, get() will still lookup the requested key in the L1 and L2 caches and return it if found. In the case when no value is found in the cache and no callback is provided, get() will return nil, nil, -1, where -1 signifies a cache miss (no value). This is not to be confused with return values such as nil, nil, 1, where 1 signifies a negative cached item found in L1 (cached nil).

Not providing a callback function allows implementing cache lookup patterns that are guaranteed to be on-cpu for a more constant, smoother latency tail end (e.g. with values refreshed in background timers via set()).

local value, err, hit_lvl = cache:get("key")
if value == nil then
    if err ~= nil then
        -- error
    elseif hit_lvl == -1 then
        -- miss (no value)
    else
        -- negative hit (cached `nil` value)
    end
end

When provided a callback, get() follows the below logic:

  1. query the L1 cache (lua-resty-lrucache instance). This cache lives in the Lua VM, and as such, it is the most efficient one to query.
    1. if the L1 cache has the value, return it.
    2. if the L1 cache does not have the value (L1 miss), continue.
  2. query the L2 cache (lua_shared_dict memory zone). This cache is shared by all workers, and is almost as efficient as the L1 cache. It however requires serialization of stored Lua tables.
    1. if the L2 cache has the value, return it.
      1. if l1_serializer is set, run it, and promote the resulting value in the L1 cache.
      2. if not, directly promote the value as-is in the L1 cache.
    2. if the L2 cache does not have the value (L2 miss), continue.
  3. create a [lua-resty-lock], and ensures that a single worker will run the callback (other workers trying to access the same value will wait).
  4. a single worker runs the L3 callback (e.g. performs a database query)
  5. the callback succeeds and returns a value: the value is set in the L2 cache, and then in the L1 cache (as-is by default, or as returned by l1_serializer if specified).
  6. the callback failed and returned nil, err: a. if resurrect_ttl is specified, and if the stale value is still available, resurrect it in the L2 cache and promote it to the L1. b. otherwise, get() returns nil, err.
  7. other workers that were trying to access the same value but were waiting are unlocked and read the value from the L2 cache (they do not run the L3 callback) and return it.

When not provided a callback, get() will only execute steps 1. and 2.

Here is a complete example usage:

local mlcache = require "mlcache"

local cache, err = mlcache.new("my_cache", "cache_shared_dict", {
    lru_size = 1000,
    ttl      = 3600,
    neg_ttl  = 60
})

local function fetch_user(user_id)
    local user, err = db:query_user(user_id)
    if err then
        -- in this case, get() will return `nil` + `err`
        return nil, err
    end

    return user -- table or nil
end

local user_id = 3

local user, err = cache:get("users:" .. user_id, nil, fetch_user, user_id)
if err then
    ngx.log(ngx.ERR, "could not retrieve user: ", err)
    return
end

-- `user` could be a table, but could also be `nil` (does not exist)
-- regardless, it will be cached and subsequent calls to get() will
-- return the cached value, for up to `ttl` or `neg_ttl`.
if user then
    ngx.say("user exists: ", user.name)
else
    ngx.say("user does not exists")
end

This second example is similar to the one above, but here we apply some transformation to the retrieved user record before caching it via the l1_serializer callback:

-- Our l1_serializer, called when a value is promoted from L2 to L1
--
-- Its signature receives a single argument: the item as returned from
-- an L2 hit. Therefore, this argument can never be `nil`. The result will be
-- kept in the L1 cache, but it cannot be `nil`.
--
-- This function can return `nil` and a string describing an error, which
-- will bubble up to the caller of `get()`. It also runs in protected mode
-- and will report any Lua error.
local function load_code(user_row)
    if user_row.custom_code ~= nil then
        local f, err = loadstring(user_row.raw_lua_code)
        if not f then
            -- in this case, nothing will be stored in the cache (as if the L3
            -- callback failed)
            return nil, "failed to compile custom code: " .. err
        end

        user_row.f = f
    end

    return user_row
end

local user, err = cache:get("users:" .. user_id,
                            { l1_serializer = load_code },
                            fetch_user, user_id)
if err then
     ngx.log(ngx.ERR, "could not retrieve user: ", err)
     return
end

-- now we can call a function that was already loaded once, upon entering
-- the L1 cache (Lua VM)
user.f()

get_bulk

syntax: res, err = cache:get_bulk(bulk, opts?)

Performs several get() lookups at once (in bulk). Any of these lookups requiring an L3 callback call will be executed concurrently, in a pool of ngx.thread.

The first argument bulk is a table containing n operations.

The second argument opts is optional. If provided, it must be a table holding the options for this bulk lookup. The possible options are:

  • concurrency: a number greater than 0. Specifies the number of threads that will concurrently execute the L3 callbacks for this bulk lookup. A concurrency of 3 with 6 callbacks to run means than each thread will execute 2 callbacks. A concurrency of 1 with 6 callbacks means than a single thread will execute all 6 callbacks. With a concurrency of 6 and 1 callback, a single thread will run the callback. Default: 3.

Upon success, this method returns res, a table containing the results of each lookup, and no error.

Upon failure, this method returns nil plus a string describing the error.

All lookup operations performed by this method will fully integrate into other operations being concurrently performed by other methods and Nginx workers (e.g. L1/L2 hits/misses storage, L3 callback mutex, etc...).

The bulk argument is a table that must have a particular layout (documented in the below example). It can be built manually, or via the new_bulk() helper method.

Similarly, the res table also has a particular layout of its own. It can be iterated upon manually, or via the each_bulk_res iterator helper.

Example:

local mlcache = require "mlcache"

local cache, err = mlcache.new("my_cache", "cache_shared_dict")

cache:get("key_c", nil, function() return nil end)

local res, err = cache:get_bulk({
  -- bulk layout:
  -- key     opts          L3 callback                    callback argument

    "key_a", { ttl = 60 }, function() return "hello" end, nil,
    "key_b", nil,          function() return "world" end, nil,
    "key_c", nil,          function() return "bye" end,   nil,
    n = 3 -- specify the number of operations
}, { concurrency = 3 })
if err then
     ngx.log(ngx.ERR, "could not execute bulk lookup: ", err)
     return
end

-- res layout:
-- data, "err", hit_lvl }

for i = 1, res.n, 3 do
    local data = res[i]
    local err = res[i + 1]
    local hit_lvl = res[i + 2]

    if not err then
        ngx.say("data: ", data, ", hit_lvl: ", hit_lvl)
    end
end

The above example would produce the following output:

data: hello, hit_lvl: 3
data: world, hit_lvl: 3
data: nil, hit_lvl: 1

Note that since key_c was already in the cache, the callback returning "bye" was never run, since get_bulk() retrieved the value from L1, as indicated by the hit_lvl value.

Note: unlike get(), this method only allows specifying a single argument to each lookup's callback.

new_bulk

syntax: bulk = mlcache.new_bulk(n_lookups?)

Creates a table holding lookup operations for the get_bulk() function. It is not required to use this function to construct a bulk lookup table, but it provides a nice abstraction.

The first and only argument n_lookups is optional, and if specified, is a number hinting the amount of lookups this bulk will eventually contain so that the underlying table is pre-allocated for optimization purposes.

This function returns a table bulk, which contains no lookup operations yet. Lookups are added to a bulk table by invoking bulk:add(key, opts?, cb, arg?):

local mlcache = require "mlcache"

local cache, err = mlcache.new("my_cache", "cache_shared_dict")

local bulk = mlcache.new_bulk(3)

bulk:add("key_a", { ttl = 60 }, function(n) return n * n, 42)
bulk:add("key_b", nil, function(str) return str end, "hello")
bulk:add("key_c", nil, function() return nil end)

local res, err = cache:get_bulk(bulk)

each_bulk_res

syntax: iter, res, i = mlcache.each_bulk_res(res)

Provides an abstraction to iterate over a get_bulk() res return table. It is not required to use this method to iterate over a res table, but it provides a nice abstraction.

This method can be invoked as a Lua iterator:

local mlcache = require "mlcache"

local cache, err = mlcache.new("my_cache", "cache_shared_dict")

local res, err = cache:get_bulk(bulk)

for i, data, err, hit_lvl in mlcache.each_bulk_res(res) do
    if not err then
        ngx.say("lookup ", i, ": ", data)
    end
end

peek

syntax: ttl, err, value = cache:peek(key, stale?)

Peek into the L2 (lua_shared_dict) cache.

The first argument key is a string which is the key to lookup in the cache.

The second argument stale is optional. If true, then peek() will consider stale values as cached values. If not provided, peek() will consider stale values, as if they were not in the cache

This method returns nil and a string describing the error upon failure.

If there is no value for the queried key, it returns nil and no error.

If there is a value for the queried key, it returns a number indicating the remaining TTL of the cached value (in seconds) and no error. If the value for key has expired but is still in the L2 cache, returned TTL value will be negative. The remaining TTL return value will only be 0 if the queried key has an indefinite ttl (ttl=0). Otherwise, this return value may be positive (key still valid), or negative (key is stale).

The third returned value will be the cached value as stored in the L2 cache, if still available.

This method is useful when you want to determine if a value is cached. A value stored in the L2 cache is considered cached regardless of whether or not it is also set in the L1 cache of the worker. That is because the L1 cache is considered volatile (as its size unit is a number of slots), and the L2 cache is still several orders of magnitude faster than the L3 callback anyway.

As its only intent is to take a "peek" into the cache to determine its warmth for a given value, peek() does not count as a query like get(), and does not promote the value to the L1 cache.

Example:

local mlcache = require "mlcache"

local cache = mlcache.new("my_cache", "cache_shared_dict")

local ttl, err, value = cache:peek("key")
if err then
    ngx.log(ngx.ERR, "could not peek cache: ", err)
    return
end

ngx.say(ttl)   -- nil because `key` has no value yet
ngx.say(value) -- nil

-- cache the value

cache:get("key", { ttl = 5 }, function() return "some value" end)

-- wait 2 seconds

ngx.sleep(2)

local ttl, err, value = cache:peek("key")
if err then
    ngx.log(ngx.ERR, "could not peek cache: ", err)
    return
end

ngx.say(ttl)   -- 3
ngx.say(value) -- "some value"

Note: since mlcache 2.5.0, it is also possible to call get() without a callback function in order to "query" the cache. Unlike peek(), a get() call with no callback will promote the value to the L1 cache, and will not return its TTL.

set

syntax: ok, err = cache:set(key, opts?, value)

Unconditionally set a value in the L2 cache and broadcasts an event to other workers so they can refresh the value from their L1 cache.

The first argument key is a string, and is the key under which to store the value.

The second argument opts is optional, and if provided, is identical to the one of get().

The third argument value is the value to cache, similar to the return value of the L3 callback. Just like the callback's return value, it must be a Lua scalar, a table, or nil. If a l1_serializer is provided either from the constructor or in the opts argument, it will be called with value if value is not nil.

On success, the first return value will be true.

On failure, this method returns nil and a string describing the error.

Note: by its nature, set() requires that other instances of mlcache (from other workers) refresh their L1 cache. If set() is called from a single worker, other workers' mlcache instances bearing the same name must call update() before their cache be requested during the next request, to make sure they refreshed their L1 cache.

Note bis: It is generally considered inefficient to call set() on a hot code path (such as in a request being served by OpenResty). Instead, one should rely on get() and its built-in mutex in the L3 callback. set() is better suited when called occasionally from a single worker, for example upon a particular event that triggers a cached value to be updated. Once set() updates the L2 cache with the fresh value, other workers will rely on update() to poll the invalidation event and invalidate their L1 cache, which will make them fetch the (fresh) value in L2.

See: update()

delete

syntax: ok, err = cache:delete(key)

Delete a value in the L2 cache and publish an event to other workers so they can evict the value from their L1 cache.

The first and only argument key is the string at which the value is stored.

On success, the first return value will be true.

On failure, this method returns nil and a string describing the error.

Note: by its nature, delete() requires that other instances of mlcache (from other workers) refresh their L1 cache. If delete() is called from a single worker, other workers' mlcache instances bearing the same name must call update() before their cache be requested during the next request, to make sure they refreshed their L1 cache.

See: update()

purge

syntax: ok, err = cache:purge(flush_expired?)

Purge the content of the cache, in both the L1 and L2 levels. Then publishes an event to other workers so they can purge their L1 cache as well.

This method recycles the lua-resty-lrucache instance, and calls ngx.shared.DICT:flush_all , so it can be rather expensive.

The first and only argument flush_expired is optional, but if given true, this method will also call ngx.shared.DICT:flush_expired (with no arguments). This is useful to release memory claimed by the L2 (shm) cache if needed.

On success, the first return value will be true.

On failure, this method returns nil and a string describing the error.

Note: it is not possible to call purge() when using a custom LRU cache in OpenResty 1.13.6.1 and below. This limitation does not apply for OpenResty 1.13.6.2 and above.

Note: by its nature, purge() requires that other instances of mlcache (from other workers) refresh their L1 cache. If purge() is called from a single worker, other workers' mlcache instances bearing the same name must call update() before their cache be requested during the next request, to make sure they refreshed their L1 cache.

See: update()

update

syntax: ok, err = cache:update(timeout?)

Poll and execute pending cache invalidation events published by other workers.

The set(), delete(), and purge() methods require that other instances of mlcache (from other workers) refresh their L1 cache. Since OpenResty currently has no built-in mechanism for inter-worker communication, this module bundles an "off-the-shelf" IPC library to propagate inter-worker events. If the bundled IPC library is used, the lua_shared_dict specified in the ipc_shm option must not be used by other actors than mlcache itself.

This method allows a worker to update its L1 cache (by purging values considered stale due to an other worker calling set(), delete(), or purge()) before processing a request.

This method accepts a timeout argument whose unit is seconds and which defaults to 0.3 (300ms). The update operation will timeout if it isn't done when this threshold in reached. This avoids update() from staying on the CPU too long in case there are too many events to process. In an eventually consistent system, additional events can wait for the next call to be processed.

A typical design pattern is to call update() only once before each request processing. This allows your hot code paths to perform a single shm access in the best case scenario: no invalidation events were received, all get() calls will hit in the L1 cache. Only on a worst case scenario (n values were evicted by another worker) will get() access the L2 or L3 cache n times. Subsequent requests will then hit the best case scenario again, because get() populated the L1 cache.

For example, if your workers make use of set(), delete(), or purge() anywhere in your application, call update() at the entrance of your hot code path, before using get():

http {
    listen 9000;

    location / {
        content_by_lua_block {
            local cache = ... -- retrieve mlcache instance

            -- make sure L1 cache is evicted of stale values
            -- before calling get()
            local ok, err = cache:update()
            if not ok then
                ngx.log(ngx.ERR, "failed to poll eviction events: ", err)
                -- /!\ we might get stale data from get()
            end

            -- L1/L2/L3 lookup (best case: L1)
            local value, err = cache:get("key_1", nil, cb1)

            -- L1/L2/L3 lookup (best case: L1)
            local other_value, err = cache:get(key_2", nil, cb2)

            -- value and other_value are up-to-date because:
            -- either they were not stale and directly came from L1 (best case scenario)
            -- either they were stale and evicted from L1, and came from L2
            -- either they were not in L1 nor L2, and came from L3 (worst case scenario)
        }
    }

    location /delete {
        content_by_lua_block {
            local cache = ... -- retrieve mlcache instance

            -- delete some value
            local ok, err = cache:delete("key_1")
            if not ok then
                ngx.log(ngx.ERR, "failed to delete value from cache: ", err)
                return ngx.exit(500)
            end

            ngx.exit(204)
        }
    }

    location /set {
        content_by_lua_block {
            local cache = ... -- retrieve mlcache instance

            -- update some value
            local ok, err = cache:set("key_1", nil, 123)
            if not ok then
                ngx.log(ngx.ERR, "failed to set value in cache: ", err)
                return ngx.exit(500)
            end

            ngx.exit(200)
        }
    }
}

Note: you do not need to call update() to refresh your workers if they never call set(), delete(), or purge(). When workers only rely on get(), values expire naturally from the L1/L2 caches according to their TTL.

Note bis: this library was built with the intent to use a better solution for inter-worker communication as soon as one emerges. In future versions of this library, if an IPC library can avoid the polling approach, so will this library. update() is only a necessary evil due to today's Nginx/OpenResty "limitations". You can however use your own IPC library by use of the opts.ipc option when creating your mlcache instance.

Resources

In November 2018, this library was presented at OpenResty Con in Hangzhou, China.

The slides and a recording of the talk (about 40 min long) can be viewed [here][talk].

Changelog

See CHANGELOG.md.

GitHub

You may find additional configuration tips and documentation for this module in the GitHub repository for nginx-module-mlcache.