Fix an unstable output issue of periodic sequence due to precision error.

Author: shunping

sdks/python/apache_beam/ml/ts/util.py

@@ -159,7 +159,19 @@ def __init__(
     self._data = data
     self._interval = interval
     self._repeat = repeat
-    self._duration = len(self._data) * interval
+
+    # In `ImpulseSeqGenRestrictionProvider`, the total number of counts
+    # (i.e. total_outputs) is computed by ceil((end - start) / interval),
+    # where end is start + duration * interval.
+    # Due to precision error of arithmetic operations, even if duration is set
+    # to len(self._data), (end - start) / interval could be a little bit smaller
+    # or bigger than len(self._data).
+    # In case of being bigger, total_outputs would be len(self._data) + 1,
+    # as the ceil() operation is used.
+    # Assuming that the precision error is no bigger than 1%, by subtracting
+    # a small amount, we ensure that the result after ceil is stable even if
+    # the precision error is present.
+    self._duration = len(self._data) * interval - 0.01 * interval
     self._max_duration = max_duration if max_duration is not None else float(
         "inf")
 

sdks/python/apache_beam/ml/ts/util_test.py

@@ -76,6 +76,17 @@ def test_timestamped_value(self):
     self.assertGreaterEqual(end - start, 3)
     self.assertLessEqual(end - start, 7)
 
+  def test_stable_output(self):
+    options = PipelineOptions()
+    data = [(Timestamp(1), 1), (Timestamp(2), 2), (Timestamp(3), 3),
+            (Timestamp(6), 6), (Timestamp(4), 4), (Timestamp(5), 5),
+            (Timestamp(7), 7), (Timestamp(8), 8), (Timestamp(9), 9),
+            (Timestamp(10), 10)]
+    expected = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
+    with beam.Pipeline(options=options) as p:
+      ret = (p | PeriodicStream(data, interval=0.0001))
+      assert_that(ret, equal_to(expected))
+
 
 if __name__ == '__main__':
   logging.getLogger().setLevel(logging.WARNING)