Plugin Development - Caching custom entities



Your plugin may need to frequently access custom entities (explained in the previous chapter) on every request and/or response. Usually, loading them once and caching them in-memory dramatically improves the performance while making sure the datastore is not stressed with an increased load.

Think of an api-key authentication plugin that needs to validate the api-key on every request, thus loading the custom credential object from the datastore on every request. When the client provides an api-key along with the request, normally you would query the datastore to check if that key exists, and then either block the request or retrieve the Consumer ID to identify the user. This would happen on every request, and it would be very inefficient:

  • Querying the datastore adds latency on every request, making the request processing slower.
  • The datastore would also be affected by an increase of load, potentially crashing or slowing down, which in turn would affect every Kong node.

To avoid querying the datastore every time, we can cache custom entities in-memory on the node, so that frequent entity lookups don't trigger a datastore query every time (only the first time), but happen in-memory, which is much faster and reliable that querying it from the datastore (especially under heavy load).

Caching custom entities

Once you have defined your custom entities, you can cache them in-memory in your code by using the singletons.cache module provided by Kong:

local singletons = require "kong.singletons"
local cache = singletons.cache

There are 2 levels of cache:

  1. L1: Lua memory cache - local to an nginx worker This can hold any type of Lua value.
  2. L2: Shared memory cache (SHM) - local to an nginx node, but shared between all workers. This can only hold scalar values, and hence requires (de)serialization.

When data is fetched from the database, it will be stored in both caches. Now if the same worker process requests the data again, it will retrieve the previously-deserialized data from the Lua memory cache. If a different worker within the same nginx node requests that data, it will find the data in the SHM, deserialize it (and store it in its own Lua memory cache) and then return it.

This module exposes the following functions:

Function name Description
value, err = cache:get(key, opts?, cb, ...) Retrieves the value from the cache. If the cache does not have value (miss), invokes cb in protected mode. cb must return one (and only one) value that will be cached. It can throw errors, as those will be caught and properly logged by Kong, at the ngx.ERR level. This function does cache negative results (nil). As such, one must rely on its second argument err when checking for errors.
ttl, err, value = cache:probe(key) Checks if a value is cached. If it is, returns its remaining TTL. It not, returns nil. The value being cached can also be a negative caching. The third return value is the value being cached itself.
cache:invalidate_local(key) Evicts a value from the node's cache.
cache:invalidate(key) Evicts a value from the node's cache and propagates the eviction events to all other nodes in the cluster.
cache:purge() Evicts all values from the node's cache.

Bringing back our authentication plugin example, to lookup a credential with a specific api-key, we would write something like:

-- access.lua
local singletons = require "kong.singletons"

local function load_entity_key(api_key)
  -- IMPORTANT: the callback is executed inside a lock, hence we cannot terminate
  -- a request here, we MUST always return.
  local apikeys, err = dao.apikeys:find_all({key = api_key}) -- Lookup in the datastore
  if err then
    error(err) -- caught by kong.cache and logged

  if not apikeys then
    return nil -- nothing was found (cached for `neg_ttl`)

  -- assuming the key was unique, we always only have 1 value...
  return apikeys[1] -- cache the credential (cached for `ttl`)

-- retrieve the apikey from the request querystring
local apikey = request.get_uri_args().apikey

-- We are using cache.get to first check if the apikey has been already
-- stored into the in-memory cache at the key: "apikeys." .. apikey
-- If it's not, then we lookup the datastore and return the credential
-- object. Internally cache.get will save the value in-memory, and then
-- return the credential.
local credential, err = cache:get("apikeys." .. apikey, nil,
                                  load_entity_key, apikey)
if err then
  return response.HTTP_INTERNAL_SERVER_ERROR(err)

if not credential then
  -- no credentials in cache nor datastore
  return responses.send_HTTP_FORBIDDEN("Invalid authentication credentials")

-- set an upstream header if the credential exists and is valid
ngx.req.set_header("X-API-Key", credential.apikey)

With the above mechanism in place, once a Consumer has made a request with their API key, the cache will be considered warm and subsequent requests won't result in a database query.

Updating or deleting a custom entity

Every time a cached custom entity is updated or deleted in the datastore (i.e. using the Admin API), it creates an inconsistency between the data in the datastore, and the data cached in the Kong nodes' memory. To avoid this inconsistency, we need to evict the cached entity from the in-memory store and force Kong to request it again from the datastore. We refer to this process as cache invalidation.

Cache invalidation for your entities

If you wish that your cached entities be invalidated upon a CRUD operation rather than having to wait for them to reach their TTL, you have to follow a few steps. This process can be automated since 0.11.0 for most entities, but manually subscribing to some CRUD events might be required to invalidate some entities with more complex relationships.

Automatic cache invalidation

Cache invalidation can be provided out of the box for your entities if you rely on the cache_key property of your entity's schema. For example, in the following schema:

local SCHEMA = {
  primary_key = { "id" },
  table = "keyauth_credentials",
  cache_key = { "key" }, -- cache key for this entity
  fields = {
    id = { type = "id" },
    created_at = { type = "timestamp", immutable = true },
    consumer_id = { type = "id", required = true, foreign = "consumers:id"},
    key = { type = "string", required = false, unique = true }

return { keyauth_credentials = SCHEMA }

We can see that we declare the cache key of this API key entity to be its key attribute. We use key here because it has a unique constraints applied to it. Hence, the attributes added to cache_key should result in a unique combination, so that no two entities could yield the same cache key.

Adding this value allows you to use the following function on the DAO of that entity:

cache_key = dao:cache_key(arg1, arg2, arg3, ...)

Where the arguments must be the attributes specified in your schema's cache_key property, in the order they were specified. This function then computes a string value cache_key that is ensured to be unique.

For example, if we were to generate the cache_key of an API key:

local cache_key = singletons.dao.keyauth_credentials:cache_key("abcd")

This would produce a cache_key for the API key "abcd" (retrieved from one of the query's arguments) that we can the use to retrieve the key from the cache (or fetch from the database if the cache is a miss):

local apikey = request.get_uri_args().apikey
local cache_key = singletons.dao.keyauth_credentials:cache_key(apikey)

local credential, err = cache:get(cache_key, nil, load_entity_key, apikey)
if err then
  return response.HTTP_INTERNAL_SERVER_ERROR(err)

-- do something with the credential

If the cache_key is generated like so and specified in an entity's schema, cache invalidation will be an automatic process: every CRUD operation that affects this API key will be make Kong generate the affected cache_key, and broadcast it to all of the other nodes on the cluster so they can evict that particular value from their cache, and fetch the fresh value from the datastore on the next request.

When a parent entity is receiving a CRUD operation (e.g. the Consumer owning this API key, as per our schema's consumer_id attribute), Kong performs the cache invalidation mechanism for both the parent and the child entity.

Note: Be aware of the negative caching that Kong provides. In the above example, if there is no API key in the datastore for a given key, the cache module will store the miss just as if it was a hit. This means that a "Create" event (one that would create an API key with this given key) is also propagated by Kong so that all nodes that stored the miss can evict it, and properly fetch the newly created API key from the datastore.

See the Clustering Guide to ensure that you have properly configured your cluster for such invalidation events.

Manual cache invalidation

In some cases, the cache_key property of an entity's schema is not flexible enough, and one must manually invalidate its cache. Reasons for this could be that the plugin is not defining a relationship with another entity via the traditional foreign = "parent_entity:parent_attribute" syntax, or because it is not using the cache_key method from its DAO, or even because it is somehow abusing the caching mechanism.

In those cases, you can manually setup your own subscriber to the same invalidation channels Kong is listening to, and perform your own, custom invalidation work. This process is similar to the old hooks.lua module that was required in < 0.11.0 for cache invalidations.

To listen on invalidation channels inside of Kong, implement the following in your plugin's init_worker handler:

function MyCustomHandler:init_worker()
  local worker_events = singletons.worker_events

  -- listen to all CRUD operations made on Consumers

  end, "crud", "consumers")

  -- or, listen to a specific CRUD operation only
    print(data.operation)  -- "update"
    print(data.old_entity) -- old entity table (only for "update")
    print(data.entity)     -- new entity table
    print(data.schema)     -- entity's schema
  end, "crud", "consumers:update")

Once the above listeners are in place for the desired entities, you can perform manual invalidations of any entity that your plugin has cached as you wish so. For instance:

  if data.operation == "delete" then
    local cache_key =
    singletons.cache:invalidate("prefix:" .. cache_key)
end, "crud", "consumers")

Extending the Admin API

As you are probably aware, the Admin API is where Kong users communicate with Kong to setup their APIs and plugins. It is likely that they also need to be able to interact with the custom entities you implemented for your plugin (for example, creating and deleting API keys). The way you would do this is by extending the Admin API, which we will detail in the next chapter: Extending the Admin API.

Next: Extending the Admin API ›

Keep up with the latest features