Update C++ code to use TemporalPolicy instead of a timestamp (#356)

Author: jparismorgan

apis/python/src/tiledb/vector_search/module.cc

@@ -17,6 +17,7 @@
 #include "detail/linalg/partitioned_matrix.h"
 #include "detail/linalg/tdb_matrix.h"
 #include "detail/linalg/tdb_partitioned_matrix.h"
+#include "detail/time/temporal_policy.h"
 
 namespace py = pybind11;
 using Ctx = tiledb::Context;
@@ -275,9 +276,8 @@ static void declare_nuv_query_heap_finite_ram(
                 centroids, query_vectors, nprobe, nthreads);
 
         auto temporal_policy{
-            (timestamp == 0) ?
-                tiledb::TemporalPolicy() :
-                tiledb::TemporalPolicy(tiledb::TimeTravel, timestamp)};
+            (timestamp == 0) ? TemporalPolicy() :
+                               TemporalPolicy(TimeTravel, timestamp)};
 
         auto mat = tdbColMajorPartitionedMatrix<T, Id_Type, Id_Type>(
             ctx,
@@ -528,9 +528,8 @@ static void declare_dist_qv(py::module& m, const std::string& suffix) {
           size_t upper_bound{0};
           auto nthreads = std::thread::hardware_concurrency();
           auto temporal_policy{
-              (timestamp == 0) ?
-                  tiledb::TemporalPolicy() :
-                  tiledb::TemporalPolicy(tiledb::TimeTravel, timestamp)};
+              (timestamp == 0) ? TemporalPolicy() :
+                                 TemporalPolicy(TimeTravel, timestamp)};
 
           py::buffer_info buf_info = active_queries_arr.request();
           auto shape = active_queries_arr.shape();
@@ -644,7 +643,11 @@ PYBIND11_MODULE(_tiledbvspy, m) {
          size_t start_pos,
          size_t end_pos,
          uint64_t timestamp) -> std::vector<uint32_t> {
-        auto r = read_vector<uint32_t>(ctx, uri, start_pos, end_pos, timestamp);
+        TemporalPolicy temporal_policy =
+            (timestamp == 0) ? TemporalPolicy() :
+                               TemporalPolicy(TimeTravel, timestamp);
+        auto r = read_vector<uint32_t>(
+            ctx, uri, start_pos, end_pos, temporal_policy);
         return r;
       });
   m.def(
@@ -654,7 +657,11 @@ PYBIND11_MODULE(_tiledbvspy, m) {
          size_t start_pos,
          size_t end_pos,
          uint64_t timestamp) -> std::vector<uint64_t> {
-        auto r = read_vector<uint64_t>(ctx, uri, start_pos, end_pos, timestamp);
+        TemporalPolicy temporal_policy =
+            (timestamp == 0) ? TemporalPolicy() :
+                               TemporalPolicy(TimeTravel, timestamp);
+        auto r = read_vector<uint64_t>(
+            ctx, uri, start_pos, end_pos, temporal_policy);
         return r;
       });
 

apis/python/src/tiledb/vector_search/type_erased_module.cc

@@ -35,14 +35,13 @@
 #include <pybind11/pybind11.h>
 #include <pybind11/stl.h>
 
+#include "api/api_defs.h"
 #include "api/feature_vector.h"
 #include "api/feature_vector_array.h"
-
 #include "api/flat_l2_index.h"
 #include "api/ivf_flat_index.h"
 #include "api/vamana_index.h"
-
-#include "api/api_defs.h"
+#include "detail/time/temporal_policy.h"
 
 namespace py = pybind11;
 
@@ -180,8 +179,13 @@ void init_type_erased_module(py::module_& m) {
              const std::string& ids_uri,
              size_t num_vectors,
              size_t timestamp) {
-            new (&instance)
-                FeatureVectorArray(ctx, uri, ids_uri, num_vectors, timestamp);
+            new (&instance) FeatureVectorArray(
+                ctx,
+                uri,
+                ids_uri,
+                num_vectors,
+                timestamp == 0 ? TemporalPolicy() :
+                                 TemporalPolicy(TimeTravel, timestamp));
           },
           py::keep_alive<1, 2>(),  // FeatureVectorArray should keep ctx alive.
           py::arg("ctx"),
@@ -276,7 +280,11 @@ void init_type_erased_module(py::module_& m) {
              const tiledb::Context& ctx,
              const std::string& group_uri,
              size_t timestamp) {
-            new (&instance) IndexVamana(ctx, group_uri, timestamp);
+            new (&instance) IndexVamana(
+                ctx,
+                group_uri,
+                timestamp == 0 ? TemporalPolicy() :
+                                 TemporalPolicy(TimeTravel, timestamp));
           },
           py::keep_alive<1, 2>(),  // IndexVamana should keep ctx alive.
           py::arg("ctx"),

apis/python/test/test_ingestion.py

@@ -512,6 +512,18 @@ def test_ingestion_with_updates(tmp_path):
             source_uri=os.path.join(dataset_dir, "data.u8bin"),
             partitions=partitions,
         )
+
+        # TODO(paris): Fix Vamana to have same metadata as Python and re-enable.
+        # with tiledb.Group(index_uri, "r", ctx={}) as group:
+        #     ingestion_timestamps = [int(x) for x in list(json.loads(group.meta.get("ingestion_timestamps", "[]")))]
+        #     base_sizes = [int(x) for x in list(json.loads(group.meta.get("base_sizes", "[]")))]
+        #     assert len(ingestion_timestamps) == 1
+        #     assert len(base_sizes) == 1
+        #     assert base_sizes[0] == 1000
+        #     timestamp_2030 = 1903946089000
+        #     timestamp_5_minutes_ago = int((time.time() - 5 * 60) * 1000)
+        #     assert ingestion_timestamps[0] > timestamp_5_minutes_ago and ingestion_timestamps[0] < timestamp_2030
+
         _, result = index.query(queries, k=k, nprobe=nprobe)
         assert accuracy(result, gt_i) == 1.0