defmodule Nebulex.Adapters.Replicated do
@moduledoc ~S"""
Built-in adapter for replicated cache topology.
## Overall features
* Replicated cache topology.
* Configurable primary storage adapter.
* Cache-level locking when deleting all entries or adding new nodes.
* Key-level (or entry-level) locking for key-based write-like operations.
* Support for transactions via Erlang global name registration facility.
* Stats support rely on the primary storage adapter.
## Replicated Cache Topology
A replicated cache is a clustered, fault tolerant cache where data is fully
replicated to every member in the cluster. This cache offers the fastest read
performance with linear performance scalability for reads but poor scalability
for writes (as writes must be processed by every member in the cluster).
Because data is replicated to all servers, adding servers does not increase
aggregate cache capacity.
There are several challenges to building a reliably replicated cache. The
first is how to get it to scale and perform well. Updates to the cache have
to be sent to all cluster nodes, and all cluster nodes have to end up with
the same data, even if multiple updates to the same piece of data occur at
the same time. Also, if a cluster node requests a lock, ideally it should
not have to get all cluster nodes to agree on the lock or at least do it in
a very efficient way (`:global` is used here), otherwise it will scale
extremely poorly; yet in the case of a cluster node failure, all of the data
and lock information must be kept safely.
The best part of a replicated cache is its access speed. Since the data is
replicated to each cluster node, it is available for use without any waiting.
This is referred to as "zero latency access," and is perfect for situations
in which an application requires the highest possible speed in its data
access.
However, there are some limitations:
* _**Cost Per Update**_ - Updating a replicated cache requires pushing
the new version of the data to all other cluster members, which will
limit scalability if there is a high frequency of updates per member.
* _**Cost Per Entry**_ - The data is replicated to every cluster member,
so Memory Heap space is used on each member, which will impact
performance for large caches.
> Based on **"Distributed Caching Essential Lessons"** by **Cameron Purdy**.
## Usage
When used, the Cache expects the `:otp_app` and `:adapter` as options.
The `:otp_app` should point to an OTP application that has the cache
configuration. For example:
defmodule MyApp.ReplicatedCache do
use Nebulex.Cache,
otp_app: :my_app,
adapter: Nebulex.Adapters.Replicated
end
Optionally, you can configure the desired primary storage adapter with the
option `:primary_storage_adapter`; defaults to `Nebulex.Adapters.Local`.
defmodule MyApp.ReplicatedCache do
use Nebulex.Cache,
otp_app: :my_app,
adapter: Nebulex.Adapters.Replicated,
primary_storage_adapter: Nebulex.Adapters.Local
end
The configuration for the cache must be in your application environment,
usually defined in your `config/config.exs`:
config :my_app, MyApp.ReplicatedCache,
primary: [
gc_interval: 3_600_000,
backend: :shards
]
If your application was generated with a supervisor (by passing `--sup`
to `mix new`) you will have a `lib/my_app/application.ex` file containing
the application start callback that defines and starts your supervisor.
You just need to edit the `start/2` function to start the cache as a
supervisor on your application's supervisor:
def start(_type, _args) do
children = [
{MyApp.ReplicatedCache, []},
...
]
See `Nebulex.Cache` for more information.
## Options
This adapter supports the following options and all of them can be given via
the cache configuration:
* `:primary` - The options that will be passed to the adapter associated
with the local primary storage. These options will depend on the local
adapter to use.
* `:task_supervisor_opts` - Start-time options passed to
`Task.Supervisor.start_link/1` when the adapter is initialized.
## Shared options
Almost all of the cache functions outlined in `Nebulex.Cache` module
accept the following options:
* `:timeout` - The time-out value in milliseconds for the command that
will be executed. If the timeout is exceeded, then the current process
will exit. For executing a command on remote nodes, this adapter uses
`Task.await/2` internally for receiving the result, so this option tells
how much time the adapter should wait for it. If the timeout is exceeded,
the task is shut down but the current process doesn't exit, only the
result associated with that task is skipped in the reduce phase.
## Telemetry events
This adapter emits all recommended Telemetry events, and documented
in `Nebulex.Cache` module (see **"Adapter-specific events"** section).
Since the replicated adapter depends on the configured primary storage
adapter (local cache adapter), this one may also emit Telemetry events.
Therefore, there will be events emitted by the replicated adapter as well
as the primary storage adapter. For example, for the cache defined before
`MyApp.ReplicatedCache`, these would be the emitted events:
* `[:my_app, :replicated_cache, :command, :start]`
* `[:my_app, :replicated_cache, :primary, :command, :start]`
* `[:my_app, :replicated_cache, :command, :stop]`
* `[:my_app, :replicated_cache, :primary, :command, :stop]`
* `[:my_app, :replicated_cache, :command, :exception]`
* `[:my_app, :replicated_cache, :primary, :command, :exception]`
As you may notice, the telemetry prefix by default for the replicated cache
is `[:my_app, :replicated_cache]`, and the prefix for its primary storage
`[:my_app, :replicated_cache, :primary]`.
See also the [Telemetry guide](http://hexdocs.pm/nebulex/telemetry.html)
for more information and examples.
## Stats
This adapter depends on the primary storage adapter for the stats support.
Therefore, it is important to ensure the underlying primary storage adapter
does support stats, otherwise, you may get unexpected errors.
## Extended API
This adapter provides some additional convenience functions to the
`Nebulex.Cache` API.
Retrieving the primary storage or local cache module:
MyCache.__primary__()
Retrieving the cluster nodes associated with the given cache name:
MyCache.nodes()
Joining the cache to the cluster:
MyCache.join_cluster()
Leaving the cluster (removes the cache from the cluster):
MyCache.leave_cluster()
## Adapter-specific telemetry events
This adapter exposes following Telemetry events:
* `telemetry_prefix ++ [:replication]` - Dispatched by the adapter
when a replication error occurs due to a write-like operation
under-the-hood.
* Measurements: `%{rpc_errors: non_neg_integer}`
* Metadata:
```
%{
adapter_meta: %{optional(atom) => term},
rpc_errors: [{node, error :: term}]
}
```
* `telemetry_prefix ++ [:bootstrap]` - Dispatched by the adapter at start
time when there are errors while synching up with the cluster nodes.
* Measurements:
```
%{
failed_nodes: non_neg_integer,
remote_errors: non_neg_integer
}
```
* Metadata:
```
%{
adapter_meta: %{optional(atom) => term},
failed_nodes: [node],
remote_errors: [term]
}
```
## Caveats of replicated adapter
As it is explained in the beginning, a replicated topology not only brings
with advantages (mostly for reads) but also with some limitations and
challenges.
This adapter uses global locks (via `:global`) for all operation that modify
or alter the cache somehow to ensure as much consistency as possible across
all members of the cluster. These locks may be per key or for the entire cache
depending on the operation taking place. For that reason, it is very important
to be aware about those operation that can potentally lead to performance and
scalability issues, so that you can do a better usage of the replicated
adapter. The following is with the operations and aspects you should pay
attention to:
* Starting and joining a new replicated node to the cluster is the most
expensive action, because all write-like operations across all members of
the cluster are blocked until the new node completes the synchronization
process, which involves copying cached data from any of the existing
cluster nodes into the new node, and this could be very expensive
depending on the number of caches entries. For that reason, adding new
nodes is considered an expensive operation that should happen only from
time to time.
* Deleting all entries. When `c:Nebulex.Cache.delete_all/2` action is
executed, like in the previous case, all write-like operations in all
members of the cluster are blocked until the deletion action is completed
(this implies deleting all cached data from all cluster nodes). Therefore,
deleting all entries from cache is also considered an expensive operation
that should happen only from time to time.
* Write-like operations based on a key only block operations related to
that key across all members of the cluster. This is not as critical as
the previous two cases but it is something to keep in mind anyway because
if there is a highly demanded key in terms of writes, that could be also
a potential bottleneck.
Summing up, the replicated cache topology along with this adapter should
be used mainly when the the reads clearly dominate over the writes (e.g.:
Reads 80% and Writes 20% or less). Besides, operations like deleting all
entries from cache or adding new nodes must be executed only once in a while
to avoid performance issues, since they are very expensive.
"""
# Provide Cache Implementation
@behaviour Nebulex.Adapter
@behaviour Nebulex.Adapter.Entry
@behaviour Nebulex.Adapter.Queryable
@behaviour Nebulex.Adapter.Stats
# Inherit default transaction implementation
use Nebulex.Adapter.Transaction
# Inherit default persistence implementation
use Nebulex.Adapter.Persistence
import Nebulex.Adapter
import Nebulex.Helpers
alias Nebulex.Cache.Cluster
alias Nebulex.{RPC, Telemetry}
## Nebulex.Adapter
@impl true
defmacro __before_compile__(env) do
otp_app = Module.get_attribute(env.module, :otp_app)
opts = Module.get_attribute(env.module, :opts)
primary = Keyword.get(opts, :primary_storage_adapter, Nebulex.Adapters.Local)
quote do
defmodule Primary do
@moduledoc """
This is the cache for the primary storage.
"""
use Nebulex.Cache,
otp_app: unquote(otp_app),
adapter: unquote(primary)
end
@doc """
A convenience function for getting the primary storage cache.
"""
def __primary__, do: Primary
@doc """
A convenience function for getting the cluster nodes.
"""
def nodes do
Cluster.get_nodes(get_dynamic_cache())
end
@doc """
A convenience function for joining the cache to the cluster.
"""
def join_cluster do
Cluster.join(get_dynamic_cache())
end
@doc """
A convenience function for removing the cache from the cluster.
"""
def leave_cluster do
Cluster.leave(get_dynamic_cache())
end
end
end
@impl true
def init(opts) do
# Required options
telemetry_prefix = Keyword.fetch!(opts, :telemetry_prefix)
telemetry = Keyword.fetch!(opts, :telemetry)
cache = Keyword.fetch!(opts, :cache)
name = opts[:name] || cache
# Maybe use stats
stats = get_boolean_option(opts, :stats)
# Primary cache options
primary_opts =
Keyword.merge(
[telemetry_prefix: telemetry_prefix ++ [:primary], telemetry: telemetry, stats: stats],
Keyword.get(opts, :primary, [])
)
# Maybe put a name to primary storage
primary_opts =
if opts[:name],
do: [name: normalize_module_name([name, Primary])] ++ primary_opts,
else: primary_opts
# Maybe task supervisor for distributed tasks
{task_sup_name, children} = sup_child_spec(name, opts)
# Prepare metadata
adapter_meta = %{
telemetry_prefix: telemetry_prefix,
telemetry: telemetry,
name: name,
primary_name: primary_opts[:name],
task_sup: task_sup_name,
stats: stats
}
# Prepare child_spec
child_spec =
Nebulex.Adapters.Supervisor.child_spec(
name: normalize_module_name([name, Supervisor]),
strategy: :rest_for_one,
children: [
{cache.__primary__, primary_opts},
{__MODULE__.Bootstrap, Map.put(adapter_meta, :cache, cache)}
| children
]
)
{:ok, child_spec, adapter_meta}
end
if Code.ensure_loaded?(:erpc) do
defp sup_child_spec(_name, _opts) do
{nil, []}
end
else
defp sup_child_spec(name, opts) do
# Task supervisor to execute parallel and/or remote commands
task_sup_name = normalize_module_name([name, TaskSupervisor])
task_sup_opts = Keyword.get(opts, :task_supervisor_opts, [])
children = [
{Task.Supervisor, [name: task_sup_name] ++ task_sup_opts}
]
{task_sup_name, children}
end
end
## Nebulex.Adapter.Entry
@impl true
defspan get(adapter_meta, key, opts) do
with_dynamic_cache(adapter_meta, :get, [key, opts])
end
@impl true
defspan get_all(adapter_meta, keys, opts) do
with_dynamic_cache(adapter_meta, :get_all, [keys, opts])
end
@impl true
defspan put(adapter_meta, key, value, _ttl, on_write, opts) do
case with_transaction(adapter_meta, on_write, [key], [key, value, opts], opts) do
:ok -> true
bool -> bool
end
end
@impl true
defspan put_all(adapter_meta, entries, _ttl, on_write, opts) do
action = if on_write == :put_new, do: :put_new_all, else: :put_all
keys = for {k, _} <- entries, do: k
with_transaction(adapter_meta, action, keys, [entries, opts], opts) || action == :put_all
end
@impl true
defspan delete(adapter_meta, key, opts) do
with_transaction(adapter_meta, :delete, [key], [key, opts], opts)
end
@impl true
defspan take(adapter_meta, key, opts) do
with_transaction(adapter_meta, :take, [key], [key, opts], opts)
end
@impl true
defspan update_counter(adapter_meta, key, amount, _ttl, _default, opts) do
with_transaction(adapter_meta, :incr, [key], [key, amount, opts], opts)
end
@impl true
defspan has_key?(adapter_meta, key) do
with_dynamic_cache(adapter_meta, :has_key?, [key])
end
@impl true
defspan ttl(adapter_meta, key) do
with_dynamic_cache(adapter_meta, :ttl, [key])
end
@impl true
defspan expire(adapter_meta, key, ttl) do
with_transaction(adapter_meta, :expire, [key], [key, ttl])
end
@impl true
defspan touch(adapter_meta, key) do
with_transaction(adapter_meta, :touch, [key], [key])
end
## Nebulex.Adapter.Queryable
@impl true
defspan execute(adapter_meta, operation, query, opts) do
do_execute(adapter_meta, operation, query, opts)
end
defp do_execute(%{name: name} = adapter_meta, :delete_all, query, opts) do
# It is blocked until ongoing write operations finish (if there is any).
# Similarly, while it is executed, all later write-like operations are
# blocked until it finishes.
:global.trans(
{name, self()},
fn ->
multi_call(adapter_meta, :delete_all, [query, opts], opts)
end,
Cluster.get_nodes(name)
)
end
defp do_execute(adapter_meta, operation, query, opts) do
with_dynamic_cache(adapter_meta, operation, [query, opts])
end
@impl true
defspan stream(adapter_meta, query, opts) do
with_dynamic_cache(adapter_meta, :stream, [query, opts])
end
## Nebulex.Adapter.Persistence
@impl true
defspan dump(adapter_meta, path, opts) do
super(adapter_meta, path, opts)
end
@impl true
defspan load(adapter_meta, path, opts) do
super(adapter_meta, path, opts)
end
## Nebulex.Adapter.Transaction
@impl true
defspan transaction(adapter_meta, opts, fun) do
super(adapter_meta, Keyword.put(opts, :nodes, Cluster.get_nodes(adapter_meta.name)), fun)
end
@impl true
defspan in_transaction?(adapter_meta) do
super(adapter_meta)
end
## Nebulex.Adapter.Stats
@impl true
defspan stats(adapter_meta) do
with_dynamic_cache(adapter_meta, :stats, [])
end
## Helpers
@doc """
Helper function to use dynamic cache for internal primary cache storage
when needed.
"""
def with_dynamic_cache(%{cache: cache, primary_name: nil}, action, args) do
apply(cache.__primary__, action, args)
end
def with_dynamic_cache(%{cache: cache, primary_name: primary_name}, action, args) do
cache.__primary__.with_dynamic_cache(primary_name, fn ->
apply(cache.__primary__, action, args)
end)
end
## Private Functions
defp with_transaction(
%{pid: pid, name: name} = adapter_meta,
action,
keys,
args,
opts \\ []
) do
nodes = Cluster.get_nodes(name)
# Ensure it waits until ongoing delete_all or sync operations finish,
# if there's any.
:global.trans(
{name, pid},
fn ->
# Write-like operation must be wrapped within a transaction
# to ensure proper replication
transaction(adapter_meta, [keys: keys, nodes: nodes], fn ->
multi_call(adapter_meta, action, args, opts)
end)
end,
nodes
)
end
defp multi_call(%{name: name, task_sup: task_sup} = meta, action, args, opts) do
task_sup
|> RPC.multi_call(
Cluster.get_nodes(name),
__MODULE__,
:with_dynamic_cache,
[meta, action, args],
opts
)
|> handle_rpc_multi_call(meta, action)
end
defp handle_rpc_multi_call({res, []}, _meta, _action), do: hd(res)
defp handle_rpc_multi_call({res, {:sanitized, {[], rpc_errors}}}, meta, action) do
_ = dispatch_replication_error(meta, action, rpc_errors)
hd(res)
end
defp handle_rpc_multi_call({responses, {:sanitized, {errors, rpc_errors}}}, meta, action) do
_ = dispatch_replication_error(meta, action, rpc_errors)
raise Nebulex.RPCMultiCallError, action: action, responses: responses, errors: errors
end
defp handle_rpc_multi_call({responses, errors}, meta, action) do
handle_rpc_multi_call({responses, {:sanitized, sanitize_errors(errors)}}, meta, action)
end
defp sanitize_errors(errors) do
Enum.reduce(errors, {[], []}, fn
{{:error, {:exception, %Nebulex.RegistryLookupError{} = error, _}}, node}, {acc1, acc2} ->
# The cache was not found in the node, maybe it was stopped and
# "Process Groups" is not updated yet, then ignore the error
{acc1, [{node, error} | acc2]}
{{:error, {:erpc, :noconnection}}, node}, {acc1, acc2} ->
# Remote node is down and maybe the "Process Groups" is not updated yet
{acc1, [{node, :noconnection} | acc2]}
error, {acc1, acc2} ->
{[error | acc1], acc2}
end)
end
defp dispatch_replication_error(adapter_meta, action, rpc_errors) do
if adapter_meta.telemetry or Map.get(adapter_meta, :in_span?, false) do
Telemetry.execute(
adapter_meta.telemetry_prefix ++ [:replication],
%{rpc_errors: length(rpc_errors)},
%{adapter_meta: adapter_meta, function_name: action, rpc_errors: rpc_errors}
)
end
end
end
defmodule Nebulex.Adapters.Replicated.Bootstrap do
@moduledoc false
use GenServer
import Nebulex.Helpers
alias Nebulex.{Adapter, Entry, Telemetry}
alias Nebulex.Adapters.Replicated
alias Nebulex.Cache.Cluster
# Max retries in intervals of 1 ms (5 seconds).
# If in 5 seconds the cache has not started, stop the server.
@max_retries 5000
## API
@doc false
def start_link(%{name: name} = adapter_meta) do
GenServer.start_link(
__MODULE__,
adapter_meta,
name: normalize_module_name([name, Bootstrap])
)
end
## GenServer Callbacks
@impl true
def init(adapter_meta) do
# Trap exit signals to run cleanup job
_ = Process.flag(:trap_exit, true)
# Ensure joining the cluster only when the cache supervision tree is started
:ok = Cluster.join(adapter_meta.name)
# Set a global lock to stop any write operation
# until the synchronization process finishes
:ok = lock(adapter_meta.name)
# Init retries
state = Map.put(adapter_meta, :retries, 0)
# Start bootstrap process
{:ok, state, 1}
end
@impl true
def handle_info(:timeout, %{pid: pid} = state) when is_pid(pid) do
# Start synchronization process
:ok = sync_data(state)
# Delete global lock set when the server started
:ok = unlock(state.name)
# Bootstrap process finished
{:noreply, state}
end
def handle_info(:timeout, %{name: name, retries: retries} = state)
when retries < @max_retries do
Adapter.with_meta(name, fn _adapter, adapter_meta ->
handle_info(:timeout, adapter_meta)
end)
rescue
ArgumentError -> {:noreply, %{state | retries: retries + 1}, 1}
end
def handle_info(:timeout, state) do
# coveralls-ignore-start
{:stop, :normal, state}
# coveralls-ignore-stop
end
@impl true
def terminate(_reason, state) do
# Ensure leaving the cluster when the cache stops
:ok = Cluster.leave(state.name)
end
## Helpers
defp lock(name) do
nodes = Cluster.get_nodes(name)
true = :global.set_lock({name, self()}, nodes)
:ok
end
defp unlock(name) do
nodes = Cluster.get_nodes(name)
true = :global.del_lock({name, self()}, nodes)
:ok
end
# FIXME: this is because coveralls does not mark this as covered
# coveralls-ignore-start
defp sync_data(%{name: name} = adapter_meta) do
cluster_nodes = Cluster.get_nodes(name)
case cluster_nodes -- [node()] do
[] ->
:ok
nodes ->
# Sync process:
# 1. Push a new generation on all cluster nodes to make the newer one
# empty.
# 2. Copy cached data from one of the cluster nodes; entries will be
# stremed from the older generation since the newer one should be
# empty.
# 3. Push a new generation on the current/new node to make it a mirror
# of the other cluster nodes.
# 4. Reset GC timer for ell cluster nodes (making the generation timer
# gap among cluster nodes as small as possible).
with :ok <- maybe_run_on_nodes(adapter_meta, nodes, :new_generation),
:ok <- copy_entries_from_nodes(adapter_meta, nodes),
:ok <- maybe_run_on_nodes(adapter_meta, [node()], :new_generation),
:ok <- maybe_run_on_nodes(adapter_meta, nodes, :reset_generation_timer) do
:ok
end
end
end
defp maybe_run_on_nodes(%{cache: cache} = adapter_meta, nodes, fun) do
if cache.__primary__.__adapter__() == Nebulex.Adapters.Local do
nodes
|> :rpc.multicall(Replicated, :with_dynamic_cache, [adapter_meta, fun, []])
|> handle_multicall(adapter_meta)
else
:ok
end
end
defp handle_multicall({responses, failed_nodes}, adapter_meta) do
{_ok, errors} = Enum.split_with(responses, &(&1 == :ok))
dispatch_bootstrap_error(
adapter_meta,
%{failed_nodes: length(failed_nodes), remote_errors: length(errors)},
%{failed_nodes: failed_nodes, remote_errors: errors}
)
end
defp copy_entries_from_nodes(adapter_meta, nodes) do
nodes
|> Enum.reduce_while([], &stream_entries(adapter_meta, &1, &2))
|> Enum.each(
&Replicated.with_dynamic_cache(
adapter_meta,
:put,
[&1.key, &1.value, [ttl: Entry.ttl(&1)]]
)
)
end
defp stream_entries(meta, node, acc) do
stream_fun = fn ->
meta
|> Replicated.stream(nil, return: :entry, page_size: 100)
|> Stream.filter(&(not Entry.expired?(&1)))
|> Stream.map(& &1)
|> Enum.to_list()
end
case :rpc.call(node, Kernel, :apply, [stream_fun, []]) do
{:badrpc, _} -> {:cont, acc}
entries -> {:halt, entries}
end
end
defp dispatch_bootstrap_error(adapter_meta, measurements, metadata) do
if adapter_meta.telemetry or Map.get(adapter_meta, :in_span?, false) do
Telemetry.execute(
adapter_meta.telemetry_prefix ++ [:bootstrap],
measurements,
Map.put(metadata, :adapter_meta, adapter_meta)
)
end
end
# coveralls-ignore-stop
end
