fix: ObservableCounter and ObservableUpDownCounter remove stale attribute combinations that are no longer observed

opentelemetry-sdk/CHANGELOG.md

@@ -5,6 +5,8 @@
 - Added `Resource::get_ref(&self, key: &Key) -> Option<&Value>` to allow retrieving a reference to a resource value without cloning.
 - **Breaking** Removed the following public hidden methods from the `SdkTracer` [#3227][3227]:
   - `id_generator`, `should_sample`
+- **Fix**: ObservableCounter and ObservableUpDownCounter now correctly report only data points from the current measurement cycle,
+  removing stale attribute combinations that are no longer observed. [#3213](https://github.com/open-telemetry/opentelemetry-rust/pull/3248)
 
 [3227]: https://github.com/open-telemetry/opentelemetry-rust/pull/3227
 

opentelemetry-sdk/src/metrics/internal/mod.rs

@@ -156,7 +156,8 @@ where
     }
 
     /// Iterate through all attribute sets and populate `DataPoints` in readonly mode.
-    /// This is used in Cumulative temporality mode, where [`ValueMap`] is not cleared.
+    /// This is used for synchronous instruments (Counter, Histogram, etc.) in Cumulative temporality mode,
+    /// where attribute sets persist across collection cycles and [`ValueMap`] is not cleared.
     pub(crate) fn collect_readonly<Res, MapFn>(&self, dest: &mut Vec<Res>, mut map_fn: MapFn)
     where
         MapFn: FnMut(Vec<KeyValue>, &A) -> Res,
@@ -179,7 +180,12 @@ where
     }
 
     /// Iterate through all attribute sets, populate `DataPoints` and reset.
-    /// This is used in Delta temporality mode, where [`ValueMap`] is reset after collection.
+    /// This is used for:
+    /// - Synchronous instruments in Delta temporality mode
+    /// - Asynchronous instruments (Observable) in both Delta and Cumulative temporality modes
+    ///
+    /// For asynchronous instruments, this removes stale attribute sets that were not observed
+    /// in the current callback, ensuring only currently active attributes are reported.
     pub(crate) fn collect_and_reset<Res, MapFn>(&self, dest: &mut Vec<Res>, mut map_fn: MapFn)
     where
         MapFn: FnMut(Vec<KeyValue>, A) -> Res,

opentelemetry-sdk/src/metrics/internal/precomputed_sum.rs

@@ -116,8 +116,10 @@ impl<T: Number> PrecomputedSum<T> {
         s_data.temporality = Temporality::Cumulative;
         s_data.is_monotonic = self.monotonic;
 
+        // Use collect_and_reset to remove stale attributes (not observed in current callback)
+        // For cumulative, report absolute values (no delta calculation needed)
         self.value_map
-            .collect_readonly(&mut s_data.data_points, |attributes, aggr| SumDataPoint {
+            .collect_and_reset(&mut s_data.data_points, |attributes, aggr| SumDataPoint {
                 attributes,
                 value: aggr.value.get_value(),
                 exemplars: vec![],

opentelemetry-sdk/src/metrics/mod.rs

@@ -1457,17 +1457,12 @@ mod tests {
         // Run this test with stdout enabled to see output.
         // cargo test asynchronous_instruments_cumulative_data_points_only_from_last_measurement --features=testing -- --nocapture
 
+        asynchronous_instruments_cumulative_data_points_only_from_last_measurement_helper("gauge");
         asynchronous_instruments_cumulative_data_points_only_from_last_measurement_helper(
-            "gauge", true,
-        );
-        // TODO fix: all asynchronous instruments should not emit data points if not measured
-        // but these implementations are still buggy
-        asynchronous_instruments_cumulative_data_points_only_from_last_measurement_helper(
-            "counter", false,
+            "counter",
         );
         asynchronous_instruments_cumulative_data_points_only_from_last_measurement_helper(
             "updown_counter",
-            false,
         );
     }
 
@@ -1782,7 +1777,6 @@ mod tests {
 
     fn asynchronous_instruments_cumulative_data_points_only_from_last_measurement_helper(
         instrument_name: &'static str,
-        should_not_emit: bool,
     ) {
         let mut test_context = TestContext::new(Temporality::Cumulative);
         let attributes = Arc::new([KeyValue::new("key1", "value1")]);
@@ -1844,12 +1838,7 @@ mod tests {
 
         test_context.flush_metrics();
 
-        if should_not_emit {
-            test_context.check_no_metrics();
-        } else {
-            // Test that latest export has the same data as the previous one
-            assert_correct_export(&mut test_context, instrument_name);
-        }
+        test_context.check_no_metrics();
 
         fn assert_correct_export(test_context: &mut TestContext, instrument_name: &'static str) {
             match instrument_name {

opentelemetry-sdk/src/metrics/pipeline.rs

@@ -524,9 +524,8 @@ fn aggregate_fn<T: Number>(
         }
         Aggregation::Sum => {
             let fns = match kind {
-                // TODO implement: observable instruments should not report data points on every collect
-                // from SDK: For asynchronous instruments with Delta or Cumulative aggregation temporality,
-                // MetricReader.Collect MUST only receive data points with measurements recorded since the previous collection
+                // Observable instruments use collect_and_reset to report only data points
+                // measured in the current callback, removing stale attributes
                 InstrumentKind::ObservableCounter => b.precomputed_sum(true),
                 InstrumentKind::ObservableUpDownCounter => b.precomputed_sum(false),
                 InstrumentKind::Counter | InstrumentKind::Histogram => b.sum(true),