A method caching macro for elixir using CAS on ETS.
Clone or download
Latest commit 97dcd38 Nov 6, 2018

README.md

Memoize

A memoization macro.

The application available in hex.pm.

Requirement

  • Elixir 1.4.5 or later.
  • Erlang/OTP 20 or later.

Installation

Add :memoize to your mix.exs dependencies:

defp deps do
  [{:memoize, "~> 1.2"}]
end

How to memoize

If you want to cache a function, use Memoize on the module and change def to defmemo.

for example:

defmodule Fib do
  def fibs(0), do: 0
  def fibs(1), do: 1
  def fibs(n), do: fibs(n - 1) + fibs(n - 2)
end

this code changes to:

defmodule Fib do
  use Memoize
  defmemo fibs(0), do: 0
  defmemo fibs(1), do: 1
  defmemo fibs(n), do: fibs(n - 1) + fibs(n - 2)
end

If a function defined by defmemo raises an error, the result is not cached and one of waiting processes will call the function.

Exclusive

A caching function that is defined by defmemo is never called in parallel.

defmodule Calc do
  use Memoize
  defmemo calc() do
    Process.sleep(1000)
    IO.puts "called!"
  end
end

# call `Calc.calc/0` in parallel using many processes.
for _ <- 1..10000 do
  Process.spawn(fn -> Calc.calc() end, [])
end

# but, actually `Calc.calc/0` is called only once.

Invalidate

If you want to invalidate cache, you can use Memoize.invalidate/{0-3}.

# invalidate a cached value of `Fib.fibs(0)`.
Memoize.invalidate(Fib, :fibs, [0])

# invalidate all cached values of `Fib.fibs/1`.
Memoize.invalidate(Fib, :fibs)

# invalidate all cached values of `Fib` module.
Memoize.invalidate(Fib)

# invalidate all cached values.
Memoize.invalidate()

Notice: Memoize.invalidate/{0-2}'s complexity is linear. Therefore, it takes a long time if Memoize has many cached values.

Caching Partial Arguments

If you want to cache with partial arguments, use Memoize.Cache.get_or_run/2 directly.

defmodule Converter do
  def convert(unique_key, data) do
    Memoize.Cache.get_or_run({__MODULE__, :resolve, [unique_key]}, fn ->
      do_convert(data)
    end)
  end
end

Cache Strategy

Cache strategy is a behaviour to management cached values.

By default, the caching strategy is Memoize.CacheStrategy.Default.

If you want to change the caching strategy, configure :cache_strategy in :memoize application.

config :memoize,
  cache_strategy: Memoize.CacheStrategy.Eviction

WARNING: A caching strategy module is determined at compile time. It mean you MUST recompile memoize module when you change the caching strategy.

memoize provides below caching strategies.

  • Memoize.CacheStrategy.Default
  • Memoize.CacheStrategy.Eviction

Cache Strategy - Memoize.CacheStrategy.Default

Default caching strategy. It provides only simple and fast features.

Basically, cached values are not collected automatically. To collect cached values, call invalidate/{0-4}, call garbage_collect/0 or specify :expires_in with defmemo.

Expiration

If you want to invalidate the cache after a certain period of time, you can use :expires_in.

defmodule Api do
  use Memoize
  defmemo get_config(), expires_in: 60 * 1000 do
    call_external_api()
  end
end

The cached value is invalidated in the first get_config/0 function call after expires_in milliseconds have elapsed.

To collect expired values, you can use garbage_collect/0. It collects all expired values. Its complexity is linear.

Cache Strategy - Memoize.CacheStrategy.Eviction

Memoize.CacheStrategy.Eviction is one of caching strategy. It provides many features, but slower than Memoize.CacheStrategy.Default.

The strategy is, basically, if cached memory size is exceeded max_threshold, unused cached values are collected until memory size falls below min_threshold.

To use Memoize.CacheStrategy.Eviction, configure :cache_strategy as below:

config :memoize,
  cache_strategy: Memoize.CacheStrategy.Eviction

config :memoize, Memoize.CacheStrategy.Eviction,
  min_threshold: 5_000_000,
  max_threshold: 10_000_000

Permanently

If :permanent option is specified with defmemo, the value won't be collected automatically. If you want to remove the value, call invalidate/{0-3}.

defmodule Json do
  use Memoize
  defmemo get_json(filename), permanent: true do
    filename |> File.read!() |> Poison.decode!()
  end
end

Notice the permanented value includes in used memory size. So you should adjust min_threshold value.

Expiration

If :expires_in option is specified with defmemo, the value will be collected after :expires_in milliseconds. To be exact, when the read/3 function is called with any arguments, all expired values will be collected.

defmodule Api do
  use Memoize
  defmemo get_config(), expires_in: 60 * 1000 do
    call_external_api()
  end
end

You can both specify :permanent and :expires_in. In the case, the cached value is not collected by garbage_collect/0 or memory size that exceed max_threshold, but after :expires_in milliseconds it is collected.

Cache Strategy - Your Strategy

You can customize caching strategy.

defmodule Memoize.CacheStrategy do
  @callback init() :: any
  @callback tab(any) :: atom
  @callback cache(any, any, Keyword.t) :: any
  @callback read(any, any, any) :: :ok | :retry
  @callback invalidate() :: integer
  @callback invalidate(any) :: integer
  @callback garbage_collect() :: integer
end

If you want to use a customized caching strategy, implement Memoize.CacheStrategy behaviour.

defmodule YourAwesomeApp.ExcellentCacheStrategy do
  @behaviour Memoize.CacheStrategy

  def init() do
    ...
  end

  ...
end

Then, configure :cache_strategy in :memoize application.

config :memoize,
  cache_strategy: YourAwesomeApp.ExcellentCacheStrategy

Notice tab/1, read/3, invalidate/{0-1}, garbage_collect/0 are called concurrently. cache/3 is not called concurrently, but other functions are called concurrently while cache/3 is called by a process.

init/0

When application is started, init/0 is called only once.

tab/1

To determine which ETS tab to use, Memoize calls tab/0.

cache/3

When new value is cached, cache/3 will be called. The first argument is key that is used as cache key. The second argument is value that is calculated value by cache key. The third argument is opts that is passed by defmemo.

cache/3 can return an any value that is called context. context is stored to ETS. And then, the context is passed to read/3's third argument.

read/3

When a value is looked up by a key, read/3 will be called. first and second arguments are same as cache/3. The third argument is context that is created at cache/3.

read/3 can return :retry or :ok. If :retry is returned, retry the lookup. If :ok is returned, return the value.

invalidte/{0,1}

These functions are called from Memoize.invalidate/{0-4}.

garbage_collect/0

The function is called from Memoize.garbage_collect/0.

Performance

  1. Generate 10,000 processes.
  2. Call 1 or 100 time(s) a memoized function for each processes. The argument of that function is random between 1 and 100.
  3. Run 5 times each benchmarks and calculate average.

Calling 1 time a memoized function means measurement of writing speed. Calling 100 time a memoized function means measurement of writing and reading speed.

Result:

Module 1 time 100 times Remarks
memoize(Default) 211 ms 999 ms
memoize(Eviction) 192 ms 1,267 ms
defmemo 233 ms 5,486 ms A memoized function was called multiple times.
cachex 10,234 ms 19,468 ms Using transaction to avoid that a memoized function is called multiple times.

Waiter config

Normally, waiter processes are waiting at the end of the computing process using message passing. However, As the number of waiting processes increases, memory is consumed, so we limit this number of the waiters.

Number of waiter processes receiving message passing are configured as config.exs or defmemo opts. (prior defmemo)

With config.exs:

config :memoize,
  max_waiter: 100,
  waiter_sleep_ms: 1000

With defmemo opts:

defmemo foo(), max_waiter: 100, waiter_sleep_ms: 1000 do
  ...
end
  • :max_waiters: Number of waiter processes receiving message passing. (default: 20)
  • :waiter_sleep_ms: Time to sleep when the number of waiter processes exceeds :max_waiters. (default: 200)

Internal

Memoize is using CAS (compare-and-swap) on ETS.

CAS is now available in Erlang/OTP 20.