rails_pod_kit 0.3.0 → 0.3.2

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 0dda5e5202f1bb5e509fad032de8449c5d8e9eec78c0cedcd0a59d20cc4f0d32
-  data.tar.gz: ebccb551ba5dcfd23c1dc2e668d417f59fb1d76f06522e02c352cea1f5e0dca5
+  metadata.gz: 136074682007e1265cd41d0d71fae7c1be18edc71edf881e82f14521391aa3bc
+  data.tar.gz: 8ab4e7116c77e72a60cda99934564124184aa1489f2ea04d5f49510293808d03
 SHA512:
-  metadata.gz: f6ab1b5432bf79341efabb50d0d3d1862dafa49523bbbc8f70ea718eb520d258b7c216da274f0d184dbeb1737c887ebfb19f04fbcc656e8fd39c6a0e057d6346
-  data.tar.gz: c2e8aa9a3b35a1ea4d557a47be04810c69e79de50a0297593388f90c1be48f8a144a281ad3d26a6c42c4564ba608d6ded064ba53b12f5f49279ddefa940e049c
+  metadata.gz: d73285ade30683bfee3f9a933b31d672dbf04fe2ac604270f49f5394799565f65aaa1bfd1a107af0a4f70b6798d75e5dd3ed295696c79cfc65528696732a32a5
+  data.tar.gz: aa042f91445f4bd90b6c6c6e7dbdce8b10e0433293efc9239b61a5046443a6a6967c0c7ad040cccc3b70c927016c1749c3f0d70d302581e5ec8a301684dd3c0a

data/README.md

@@ -609,9 +609,29 @@ serves `solid_queue_*` and nothing else, so the check config needs no filters.
 - **Puma control app.** `yabeda-puma-plugin` reads Puma's thread-pool stats
   through Puma's control app, so `Puma.activate` activates one on a
   localhost-only socket (`no_token: true`, never network-exposed).
+- **`json` 3 needs Rails 8.1.** On ActiveSupport **< 8.1**, json **>= 3** breaks
+  *every* `to_json` call in the process, not just this gem's. ActiveSupport up
+  to 8.0 encodes with `JSON.generate(..., quirks_mode: true)`; json 3.0 turned
+  that unknown keyword from a silent no-op into an `ArgumentError`, and Rails
+  dropped the call only in 8.1. The whole ecosystem is pinning `json` to `"< 3"`
+  behind a Rails-version guard, and so should a host on 7.2 or 8.0 — the
+  appraisals here do exactly that.
+
+  What is specific to this gem is the *shape* of its symptom. Everywhere else
+  the error is loud; in the metrics path it is not. prometheus-client-mmap
+  builds each metric's mmap key with `to_json` and catches the failure in
+  `UsesValueType#value_object`, logging at INFO before falling back to
+  `SimpleValue` — which keeps the value in the recording process instead of the
+  shared mmap file. So `/metrics` still answers 200 and the series simply go
+  missing. If you are chasing an empty exposition, check this first.
+
+  Not expressible as a gemspec dependency: it applies only below ActiveSupport
+  8.1, and `json < 3` there would hold back Rails 8.1+ hosts, which are fine
+  (verified: AS 8.1.3.1 + json 3.0.2 encodes normally).
+
 - **Rack version.** Under **Rack 3+** the mmap exporter's WEBrick handler also
   needs the `rackup` gem. Under Rack 2.x `webrick` alone is enough, but on
-  Ruby ≥ 3.5 make sure `ostruct` is in the bundle (Rack 2.2 requires it
+  Ruby ≥ 4.0 make sure `ostruct` is in the bundle (Rack 2.2 requires it
   without declaring it, and it's no longer a default gem).
 - **Two kinds of entry point.** The main file (`require 'rails_pod_kit'`,
   what Bundler.require loads in a Rails app) pulls in every integration

data/VERSION

@@ -1 +1 @@
-0.3.0
+0.3.2

data/lib/rails_pod_kit/health.rb

@@ -36,10 +36,11 @@ module RailsPodKit
     # Configures HealthMonitor with the kit's defaults.
     #
     # redis:   optional Redis connection options hash (same shape the host
-    #          passes to Sidekiq) or a ready connection object (Redis /
-    #          ConnectionPool); when given the Redis provider is added. Omit
-    #          entirely (or pass nil) on hosts with no Redis dependency to get
-    #          a database + cache only endpoint.
+    #          passes to Sidekiq) or a ready connection object (Redis,
+    #          ConnectionPool, or a ConnectionPool::Wrapper, which is unwrapped
+    #          to the pool behind it); when given the Redis provider is added.
+    #          Omit entirely (or pass nil) on hosts with no Redis dependency to
+    #          get a database + cache only endpoint.
     # path:    mount-relative endpoint path (default :healthz).
     # sidekiq: optional thresholds hash; when given the Sidekiq provider is
     #          added with the provided `queue_size:` / `latency:` overrides
@@ -88,6 +89,17 @@ module RailsPodKit
     end
 
     def build_connection(redis)
+      # ConnectionPool::Wrapper is a BasicObject delegator: it defines no
+      # predicates of its own, so every `is_a?` on it falls through to
+      # method_missing and checks out a real connection. That costs a
+      # connection attempt here at boot -- fatal when the broker is down, the
+      # very outage a health endpoint exists to report -- and health_monitor's
+      # own `is_a?(ConnectionPool)` test misses the wrapper anyway and nests it
+      # in a second pool, paying another checkout per command. Hand over the
+      # pool it wraps. `respond_to?` is answered from the wrapper's own
+      # allowlist without touching the network, and is false for the hash,
+      # Redis and ConnectionPool arguments handled below.
+      return redis.wrapped_pool if redis.respond_to?(:wrapped_pool)
       return redis unless redis.is_a?(Hash)
 
       require 'redis'

data/lib/rails_pod_kit/puma.rb

@@ -43,10 +43,16 @@ module RailsPodKit
       puma_config.plugin :yabeda
       puma_config.plugin :yabeda_prometheus
 
-      # Silence the exporter's per-scrape access log. The `prometheus_silence_logger`
-      # DSL method is defined by the :yabeda_prometheus plugin, so this must run
-      # after the plugin is loaded above. See Config#silence_exporter_access_log.
-      puma_config.prometheus_silence_logger(true) if RailsPodKit.config.silence_exporter_access_log
+      # Drop the exporter's per-scrape access log the same way the WEBrick path
+      # does (see RailsPodKit::Exporter.start!): the log line comes from the
+      # Rack::CommonLogger the exporter's rack app mounts unless this is exactly
+      # 'false'. See Config#silence_exporter_access_log.
+      #
+      # Deliberately not `prometheus_silence_logger(true)`, which is the plugin's
+      # own knob: it swaps Puma's whole log writer for LogWriter.null, and that
+      # writer also carries the exporter's errors — so a /metrics that raises on
+      # every scrape would fail completely silently.
+      ENV['PROMETHEUS_EXPORTER_LOG_REQUESTS'] = 'false' if RailsPodKit.config.silence_exporter_access_log
 
       # `config/puma.rb` is evaluated before Rails is loaded, so requiring this
       # gem here loads yabeda *before* `defined?(Rails)`, and yabeda's Railtie

data/lib/rails_pod_kit/sidekiq.rb

@@ -21,6 +21,8 @@ module RailsPodKit
     def install!(sidekiq_config = nil)
       return unless RailsPodKit.enabled?
 
+      export_policy_env!(collect_cluster_metrics: RailsPodKit.config.sidekiq_global_metrics == :all)
+
       require 'yabeda/sidekiq'
       require 'yabeda/prometheus/mmap'
 
@@ -72,9 +74,10 @@ module RailsPodKit
     # server — i.e. the web (Puma) process under the :web policy. The web has the
     # Sidekiq client configured (Redis access), so yabeda-sidekiq can read the
     # cluster stats there. We force `collect_cluster_metrics` on and keep
-    # `declare_process_metrics` off (the web runs no jobs). Must run before
-    # Yabeda.configure! so the gauges are declared.
+    # `declare_process_metrics` off (the web runs no jobs).
     def enable_global_collection!
+      export_policy_env!(collect_cluster_metrics: true, declare_process_metrics: false)
+
       require 'yabeda/sidekiq'
 
       Yabeda::Sidekiq.config.collect_cluster_metrics = true
@@ -82,6 +85,24 @@ module RailsPodKit
       apply_retries_segmentation!
     end
 
+    # Publishes the policy through yabeda-sidekiq's own env-backed config, and it
+    # has to happen *before* `require 'yabeda/sidekiq'`.
+    #
+    # yabeda-sidekiq declares its cluster gauges inside its own `Yabeda.configure`
+    # block, guarded by `collect_cluster_metrics` — and when Yabeda has already
+    # been configured (any host where its Railtie ran first) requiring the file
+    # evaluates that block immediately. Assigning the flag afterwards is then too
+    # late: the gauges are never declared, while the `collect` block — which reads
+    # the same flag on every scrape — starts referencing them, and /metrics 500s
+    # with a NameError. The env values are read when the config object is built,
+    # so setting them here makes the declaration see the policy whenever the
+    # require lands.
+    def export_policy_env!(collect_cluster_metrics:, declare_process_metrics: nil)
+      ENV['YABEDA_SIDEKIQ_COLLECT_CLUSTER_METRICS'] = collect_cluster_metrics.to_s
+      ENV['YABEDA_SIDEKIQ_DECLARE_PROCESS_METRICS'] = declare_process_metrics.to_s unless declare_process_metrics.nil?
+      ENV['YABEDA_SIDEKIQ_RETRIES_SEGMENTED_BY_QUEUE'] = RailsPodKit.config.retries_segmented_by_queue.to_s
+    end
+
     # Starts the background WEBrick exporter shared with the other non-Puma
     # entry points; guarded there so a re-entrant Sidekiq boot can't double-bind
     # the port.

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: rails_pod_kit
 version: !ruby/object:Gem::Version
-  version: 0.3.0
+  version: 0.3.2
 platform: ruby
 authors:
 - Fabio Napoleoni