Add prometheus.relabel external cache proposal

docs/design/1600-prometheus-relabel-external-cache.md

@@ -0,0 +1,172 @@
+# Proposal: Prometheus relabel external cache
+
+- Author(s): Wilfried ROSET (@wilfriedroset), Pierre BAILHACHE (@pbailhache)
+- Last updated: 2024-09-02
+- Original issue: <https://github.com/grafana/alloy/issues/1600>
+
+## Abstract
+
+This proposal introduces a way to configure the component `prometheus.relabel` so that it could use an external cache such as `Redis` or `Memcached` instead of `in-memory`.
+
+## Problem
+
+The `prometheus.relabel` component rewrites the label set of each metric passed along to the exported receiver by applying one or more relabeling rules. To do so it uses a relabeling cache stored in memory. For each `prometheus.relabel` component Alloy will create a dedicated relabel cache. This is not a huge issue per se due to the possibility of registering multiple rules into on component to mutualize the cache.
+
+However if you're horizontally scaling Alloy deployment with load-balancing, you will end up with one cache per Alloy instance and those local caches will have overlaps, increasing the footprint of each instance.
+Moreover, the cache is tied to the pod, which means that a new Alloy process starts with an empty cache.
+
+For a couple of horizontally scaled Alloy pods it's acceptable but if you plan to have lots of instances processing data horizontally it's not sustainable.
+
+## Proposal
+
+Allow to use `Redis` or `Memcached` instead of the `in-memory` one as cache.
+
+Using [dskit](https://github.com/grafana/dskit/blob/main/cache/cache.go) code to manage connection and client configuration.
+
+We could create an interface to avoid changing the code logic and to abstract the kind of cache used to the component.
+
+## Pros and cons
+
+**Pros:**
+
+- No logic change so impact is expected to be null for users
+- Possibility to use an external cache if needed, even having multiple caches for different relabeling components.
+- Will be easy to use in other parts of the codebase
+
+**Cons:**
+
+- Config is a bit more complex compared to previous one
+
+## Alternative solutions
+
+The alternative is to do nothing as we deem this improvement unnecessary.
+
+## Compatibility
+
+This proposal offer to deprecate for a couple of release the old way to configure the in-memory cache, then release+2 drop the old way. Doing so allow to migrate the settings to the correct block.
+
+## Implementation
+
+We should add a re-usable cache interface compatible with multiple backend, it should be usable with different value types
+
+```golang
+type Cache[valueType any] interface {
+ Get(key string) (*valueType, error)
+ Set(key string, value *valueType, ttl time.Duration) error
+ Remove(key string)
+ GetMultiple(keys []string) ([]*valueType, error)
+ SetMultiple(values map[string]*valueType, ttl time.Duration) error
+ Clear(newSize int) error
+ GetCacheSize() int
+}
+```
+
+Each backend cache should be implemented in a separate file. For now we should support in_memory, redis and memcached.
+
+We should add a way to select the cache and its connections options through the components arguments
+
+For example, based on what's done in [Mimir index cache](https://github.com/grafana/mimir/blob/main/pkg/storage/tsdb/index_cache.go#L47):
+
+```golang
+type Arguments struct {
+    // Where the relabelled metrics should be forwarded to.
+    ForwardTo []storage.Appendable `alloy:"forward_to,attr"`
+
+    // The relabelling rules to apply to each metric before it's forwarded.
+    MetricRelabelConfigs []*alloy_relabel.Config `alloy:"rule,block,optional"`
+
+    // DEPRECATED Use type = inmemory and cache_size field.
+    InMemoryCacheSizeDeprecated int `alloy:"max_cache_size,attr,optional"`
+
+    // Cache backend configuration.
+    CacheConfig cache.CacheConfig `alloy:"cache,block,optional"`
+}
+
+type CacheConfig struct {
+   cache.BackendConfig `yaml:",inline"`
+   InMemory            InMemoryCacheConfig `yaml:"inmemory"`
+}
+
+type InMemoryCacheConfig struct {
+   CacheSize int `yaml:"cache_size"`
+}
+
+------
+type BackendConfig struct {
+    Backend   string                `yaml:"backend"`
+    Memcached MemcachedClientConfig `yaml:"memcached"`
+    Redis     RedisClientConfig     `yaml:"redis"`
+}
+```
+
+Configuration should be `in_memory` by default.
+`max_cache_size` should still be taken into account but only if `backend = in_memory`. It also should be deprecated and we should redirect to the `InMemoryRelabelCacheConfig.CacheSize` field.
+
+Here is some examples:
+
+- legacy config unchanged
+
+```river
+prometheus.relabel "legacy_config" {
+  forward_to = [...]
+  max_cache_size = 10000000
+  rule {
+   ...
+  }
+
+}
+```
+
+- redis config
+
+```river
+prometheus.relabel "redis_config" {
+  forward_to = [...]
+  cache {
+    backend = "redis"
+    redis {
+        endpoint = "redis.url"
+        username = "user"
+        password = "password"
+        ...
+    }
+  }
+  ...
+}
+```
+
+- new in memory config
+
+```river
+prometheus.relabel "inmemory_config" {
+  forward_to = [...]
+  cache {
+    backend = "inmemory"
+    inmemory {
+        cache_size = 10000000
+    }
+  }
+  ...
+}
+```
+
+- memcached config
+
+```river
+prometheus.relabel "memcached_config" {
+  forward_to = [...]
+  cache {
+    backend = "memcached"
+    memcached {
+        addresses = "address1, address2"
+        timeout = 10
+        ...
+    }
+  }
+  ...
+}
+```
+
+## Related open issues
+
+N/A