Avoid encoding context for cached Tabular predictions

packages/structured-data-models-client/src/sdm_client/adapters/tabular.py

@@ -112,7 +112,9 @@ def session_prediction_plan(
     ) -> Iterator[dict[str, Any]]:
         declared = _declared_by(pinned)
         try:
-            envelope = self._envelope(request, outputs, declared)
+            envelope = self._envelope(
+                request, outputs, declared, include_context=False
+            )
         except Exception as error:
             # Writing the rows to the pinned schema is what makes the call
             # legible to the endpoint, and it is also the only thing that can
@@ -185,6 +187,8 @@ def _envelope(
         request: Request,
         outputs: tuple[str, ...],
         pinned_dtypes: Mapping[str, str] | None = None,
+        *,
+        include_context: bool = True,
     ) -> dict[str, Any]:
         if request.payloads is not None:
             raise UnsupportedRequestError(
@@ -284,11 +288,16 @@ def _envelope(
                 },
             },
             'schema': {'instance_table': instance_schema},
-            'context': {'instance_table': encode_table(context, dtypes)},
             'predict': {'instance_table': encode_table(predict, dtypes)},
             'output': {'fields': list(outputs)},
             'metadata': {'request_id': _request_id()},
         }
+        # Cached scoring keeps the schema checks above, but the context is
+        # already on the server. Do not encode rows that will not be sent.
+        if include_context:
+            envelope['context'] = {
+                'instance_table': encode_table(context, dtypes)
+            }
         if request.inference:
             envelope['inference'] = typed_inference(
                 json_ready(request.inference)

packages/structured-data-models-client/tests/test_envelope_and_result.py

@@ -141,6 +141,38 @@ def test_a_session_pins_the_context_and_scoring_sends_only_the_rows():
     assert 'context' not in scoring
 
 
+@pytest.mark.parametrize('task', ['classification', 'regression'])
+@pytest.mark.parametrize('with_pinned_schema', [False, True])
+def test_session_scoring_does_not_encode_the_context(
+    monkeypatch, task, with_pinned_schema
+):
+    from sdm_client.adapters import tabular
+
+    adapter = TabularAdapter()
+    request = request_for(task, inference={'num_context_samples': 500})
+    outputs = ('prediction',)
+    full = next(adapter.prediction_plan(request, outputs))
+    pinned = adapter.session_plan(request, outputs)
+    expected = {key: full[key] for key in ('predict', 'output', 'inference')}
+    encode = tabular.encode_table
+    encoded_rows = []
+
+    def encode_scoring_rows(frame, dtypes):
+        assert frame is not request.context.instance_table
+        encoded_rows.append(len(frame))
+        return encode(frame, dtypes)
+
+    monkeypatch.setattr(tabular, 'encode_table', encode_scoring_rows)
+    actual = next(
+        adapter.session_prediction_plan(
+            request, outputs, pinned if with_pinned_schema else None
+        )
+    )
+    assert encoded_rows == [len(request.predict.instance_table)]
+    assert actual.pop('metadata')['request_id'].startswith('sdm-')
+    assert actual == expected
+
+
 def test_every_request_carries_an_id_that_correlates_it_with_server_logs():
     assert envelope()['metadata']['request_id'].startswith('sdm-')