rails-cache-performance

name	rails-cache-performance
description	Rails cache invalidation, stale cache detection, request-level fragment caching, and Russian doll caching patterns.
license	MIT
metadata	{"version":"1.1.0","hermes":{"tags":["Rails","Caching","Performance","Patterns"],"related_skills":["rails-testing-patterns","activerecord-application-query-optimization"]}}

Rails Cache Performance

Fragment Cache Invalidation

Use touch cascades so parent records auto-invalidate when children update:

class Image < ApplicationRecord
  belongs_to :viewable, polymorphic: true, touch: true
end

Costs to weigh: every child save issues an extra parent UPDATE — on hot parents (many concurrent child writes) that's row-lock contention and write amplification; bulk child updates become touch storms. And the cascade is one level: Russian-doll invalidation up a deeper graph needs touch: true at EVERY intermediate belongs_to, paired with nested cache blocks in the views keyed on each record.

Multi-Layer Cache Clearing

Clear all layers: Rails.cache.delete(key) (app cache), @cached_value = nil (memoization), reload (association cache).

Stale Cache Detection

Routing caches storing record IDs cause sporadic 404s when cached IDs no longer exist. Use atomic fetch with stampede prevention:

record = Rails.cache.fetch(cache_key,
                          expires_in: 10.seconds,
                          race_condition_ttl: 5.seconds) do
  find_by(name: lookup_name)
end

Simpler alternative: for indexed lookups (<1ms), shrink TTL from minutes to 10 seconds — zero logic changes; the stale window shrinks proportionally to the TTL cut (e.g. 1-5min → 10s).

For tenant/subdomain resolution specifically, a stale cached ID is not just a 404: if the name is freed and reassigned within the TTL window, requests can resolve to the WRONG tenant's record — cross-tenant exposure. That is the concrete reason the security note below says to consider no cache at all on those paths.

Cache TTL Selection

TTL	Use Case	Recommendation
10s	Routing/resolution (subdomain, tenant)	Default for routing
30s	Stable data, higher traffic	Good middle ground
1min	Very stable data	Maximum for routing
5min+	Static reference data	Avoid for routing

For security-critical paths (multi-tenant resolution, auth): consider removing cache entirely. ~2ms per-request cost is negligible vs. data breach risk.

Request-Scoped Caching for Current.* Attributes

Check if already set before calling the setter — without early return, each partial/helper call re-evaluates:

def current_campaign
  return Current.campaign if Current.campaign.present?
  set_current_campaign
end

Nil never latches with this guard: when there legitimately IS no campaign, the setter re-runs on every call. If absent-is-common, memoize the LOOKED-UP state (a Current.campaign_resolved flag) rather than the value's presence.

Log Volume Reduction

Enable lograge, optionally suppressing ActionView render notifications (the unsubscribe lines, not lograge itself, are what removes render lines — a large reduction in render-heavy apps):

config.lograge.enabled = true
config.lograge.ignore_actions = ['HealthController#index']
ActiveSupport::Notifications.unsubscribe("render_template.action_view")
ActiveSupport::Notifications.unsubscribe("render_partial.action_view")

unsubscribe is global: it removes EVERY subscriber of that event — including APM instrumentation (New Relic, Skylight, AppSignal, Datadog view-render metrics). If you run an APM, skip the unsubscribe lines or your view breakdowns go silently blank.