atlas_rb 1.27.2 → 1.27.3

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 3bf07fbfa75e57be6651cc352fc2216f8c5e4029a68393018fa14b8e6a34e6aa
-  data.tar.gz: 63d66617e658bc0693c7a07afdf0b38cac2067f464b3ba1e0466b3a8d7d1860e
+  metadata.gz: 57da54874f04fe2d2b07757c8c062fe5a4e95aae8d843de69ed7bd53158b8122
+  data.tar.gz: af25098d8000917fac7e65121fc575dd38460ed23c2515e6309d2d230c9389ce
 SHA512:
-  metadata.gz: 2071b3face694f62a03db93362462bf77381cfdcee77c138ff5a2baabe47f000d94ede673ca1d28f160d0348506a844346f920b71457669bae326769aa9286e3
-  data.tar.gz: 823ca671c648afad8c4a82111d658a691c6bf757e02cb942594c0a1c6afec2a1b6fda628c954ca909497c7168a6384f845fcf4ed229a1f470d5fabb1bac212c8
+  metadata.gz: 228ce5a5c9c2746628bea89760fd5b623d017a3511e2d8f59ce22401aa58499a2e82500da0fd76ba12dacd5e3d5fd772694f717b7312e98b0c2b8fec43c94ca2
+  data.tar.gz: 3b92e9cf2b59e766276a7654fa2769d5339b9b511f146aa429692b8f7d84dfa3a96b5575659b6ee923d64bc3b08bc1291f8245de75922b419ba9b276629a7043

data/.version

@@ -1 +1 @@
-1.27.2
+1.27.3

data/CHANGELOG.md

@@ -1,5 +1,14 @@
 # Changelog
 
+## 1.27.3
+
+### Changed — `Metrics.latency` documentation
+
+Atlas 0.6.227 times each request from the top of its Rack stack, the same span
+as `x-runtime`, rather than the controller action alone. The YARD and README
+now say so. No code changed; against an older Atlas the figures are still
+action-only.
+
 ## 1.27.2
 
 Requires an Atlas that serves `GET /metrics/latency`.

data/Gemfile.lock

@@ -1,7 +1,7 @@
 PATH
   remote: .
   specs:
-    atlas_rb (1.27.2)
+    atlas_rb (1.27.3)
       faraday (~> 2.7)
       faraday-follow_redirects (~> 0.3.0)
       faraday-multipart (~> 1)

data/README.md

@@ -871,9 +871,11 @@ latency["p95_ms"]         # => 98
 latency["slowest_actions"].first  # => { "action" => "works#mods", "calls" => 912, "p95_ms" => 188 }
 ```
 
-Atlas times each request from when Rails dispatches it until it returns, so
-the figures leave out the network and your own waiting. If you also time your
-Atlas calls, the difference is time spent on your side. The percentiles are
+Atlas times each request from when it enters Atlas's Rack stack until the app
+returns the response, the same span as the `x-runtime` header. The figures
+include time a request waited inside Atlas behind concurrent ones. They leave
+out time queued for a Puma thread, the network, and your own waiting. If you
+also time your Atlas calls, the difference is time spent on your side. The percentiles are
 interpolated from histogram buckets, so they are approximate.
 
 Like `Health.read`, it has a 5-second deadline and no retry. It raises

data/lib/atlas_rb/metrics.rb

@@ -15,11 +15,14 @@ module AtlasRb
 
     # How long Atlas took to serve its own requests (`GET /metrics/latency`).
     #
-    # Measured inside Atlas, from when Rails dispatches the action until it
-    # returns, so it leaves out the network and the caller's own waiting. A
-    # caller that also times its Atlas calls can subtract to see how much of
-    # its wait was its own. The percentiles are interpolated from histogram
-    # buckets, so they are approximate.
+    # Measured inside Atlas, from when a request enters its Rack stack until
+    # the app returns the response — the same span as the `x-runtime` header.
+    # So it includes time a request spent waiting inside Atlas behind
+    # concurrent ones, and leaves out time queued for a Puma thread, the
+    # network, and the caller's own waiting. A caller that also times its
+    # Atlas calls can subtract to see how much of its wait was its own. The
+    # percentiles are interpolated from histogram buckets, so they are
+    # approximate.
     #
     # No retry, and a fresh socket per call: like {AtlasRb::Health.read}, it is
     # polled rarely and should fail fast rather than hide a struggling Atlas.

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: atlas_rb
 version: !ruby/object:Gem::Version
-  version: 1.27.2
+  version: 1.27.3
 platform: ruby
 authors:
 - David Cliff