[DRAFT] Add TVP Support for JSON and Vector (float32) by apoorvdeshmukh · Pull Request #4458 · dotnet/SqlClient

apoorvdeshmukh · 2026-07-20T16:18:18Z

Description

This PR is a work in progress for adding TVP support for JSON and vector(float32) datatypes.

Issues

#3449

Testing

Added functional and manual tests where applicable.

Guidelines

Please review the contribution guidelines before submitting a pull request:

Copilot

Pull request overview

This draft PR adds end-to-end Table-Valued Parameter (TVP) support for the new SQL Server json and vector(float32) types by extending the SMI/TDS metadata and value plumbing, and by adding functional/manual validation tests.

Changes:

Extend TVP TYPE_INFO serialization to support json (as NVARCHAR(max) + collation) and vector (token + byte length + element-type indicator).
Update SMI metadata/value utilities to allow JSON and Vector columns in SqlMetaData / SqlDataRecord and to map accessors/setters appropriately.
Add functional and manual tests validating JSON/Vector TVP round-trips and insert scenarios.

Reviewed changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
src/Microsoft.Data.SqlClient/tests/ManualTests/SQL/VectorTest/NativeVectorFloat32Tests.cs	Adds manual TVP round-trip and insert tests for `vector(float32)`.
src/Microsoft.Data.SqlClient/tests/ManualTests/SQL/JsonTest/JsonTest.cs	Adds manual TVP round-trip and insert tests for `json`.
src/Microsoft.Data.SqlClient/tests/FunctionalTests/SqlMetaDataTest.cs	Adds in-memory (no-server) functional coverage for JSON/Vector `SqlMetaData` + `SqlDataRecord`.
src/Microsoft.Data.SqlClient/src/Microsoft/Data/SqlClient/TdsValueSetter.cs	Enables writing NULL for non-PLP Vector columns via the varlength null marker.
src/Microsoft.Data.SqlClient/src/Microsoft/Data/SqlClient/TdsParser.cs	Serializes TVP TYPE_INFO for JSON and Vector.
src/Microsoft.Data.SqlClient/src/Microsoft/Data/SqlClient/Server/ValueUtilsSmi.cs	Adds getter/setter plumbing for JSON/Vector values and access map handling.
src/Microsoft.Data.SqlClient/src/Microsoft/Data/SqlClient/Server/SqlMetaData.cs	Adds default metadata and constructor support for JSON/Vector.
src/Microsoft.Data.SqlClient/src/Microsoft/Data/SqlClient/Server/SmiXetterAccessMap.cs	Maps JSON/Vector xetter validity to equivalent existing types (NVarChar/VarBinary).
src/Microsoft.Data.SqlClient/src/Microsoft/Data/SqlClient/Server/SmiMetaData.cs	Adds default SMI metadata and type-name entries for JSON/Vector.
src/Microsoft.Data.SqlClient/src/Microsoft/Data/SqlClient/Server/MetadataUtilsSmi.cs	Extends type-code inference and schema-row metadata building for JSON/Vector.

            DBNull.Value,       // SqlDbType.DateTime2
            DBNull.Value,       // SqlDbType.DateTimeOffset
+            SqlString.Null,     // SqlDbTypeExtensions.Json (value 35)
+            SqlBinary.Null,     // SqlDbTypeExtensions.Vector (value 36)


+            else if (SqlDbTypeExtensions.Vector == dbType)
+            {
+                // Vector max length is the total byte size (8-byte header + element payload).
+                if (maxLength < TdsEnums.VECTOR_HEADER_SIZE)
+                {
+                    throw ADP.Argument(StringsHelper.GetString(Strings.ADP_InvalidDataLength2, maxLength.ToString(CultureInfo.InvariantCulture)), nameof(maxLength));
+                }
+            }


+                        case SqlDbTypeExtensions.Json:
+                            {
+                                Debug.Assert(CanAccessSetterDirectly(metaData[i], ExtendedClrTypeCode.String));
+                                // JSON travels as a single UTF-8 PLP value (see TdsValueSetter.SetString).


+                        case SqlDbTypeExtensions.Json:
+                            {
+                                Debug.Assert(CanAccessSetterDirectly(metaData[i], ExtendedClrTypeCode.String));
+                                // JSON travels as a single UTF-8 PLP value (see TdsValueSetter.SetString).


codecov · 2026-07-20T17:53:50Z

Codecov Report

❌ Patch coverage is 83.33333% with 18 lines in your changes missing coverage. Please review.
✅ Project coverage is 65.33%. Comparing base (fdebcd2) to head (7fba6d8).
⚠️ Report is 18 commits behind head on main.

Files with missing lines	Patch %	Lines
...c/Microsoft/Data/SqlClient/Server/ValueUtilsSmi.cs	74.13%	15 Missing ⚠️
...src/Microsoft/Data/SqlClient/Server/SqlMetaData.cs	85.71%	2 Missing ⚠️
...icrosoft/Data/SqlClient/Server/MetadataUtilsSmi.cs	80.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #4458      +/-   ##
==========================================
- Coverage   65.83%   65.33%   -0.50%     
==========================================
  Files         287      285       -2     
  Lines       43763    66986   +23223     
==========================================
+ Hits        28812    43767   +14955     
- Misses      14951    23219    +8268

Flag	Coverage Δ
CI-SqlClient	`?`
PR-SqlClient-Project	`65.33% <83.33%> (?)`

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

edwardneal

Coincidentally, I'd been trying to get these types flowing as TVP fields over the last few days! I'd bumped into a problem with undocumented TYPE_INFO for vector and some transport difficulties with json.

Most of these are just questions or ad-hoc housekeeping. The two key points which I'm unsure about are the code paths used for DataTable instances, and whether we should really be transporting JSON data as Unicode-encoded nvarchars.

edwardneal · 2026-07-20T17:29:22Z

+                    WriteUnsignedInt(_defaultCollation._info, stateObj);
+                    stateObj.WriteByte(_defaultCollation._sortId);


Not all collations support every character (é is a decent example here.) In such circumstances, SQL Server will strip away the accent. If the user's database uses one of these collations, will SQL Server convert a JSON string of {"a":"café"} to {"a": "cafe"}?

edwardneal · 2026-07-20T17:32:14Z

+            // JSON and Vector are outside the contiguous SqlDbType range covered by this map;
+            // JSON accepts the same setters as NVarChar (string) and Vector the same as VarBinary (bytes).


The array runs from BigInt to DateTimeOffset. Json and Vector directly follow this. I'm not sure we need the special handling here.

There are a handful of other places which this also applies to.

edwardneal · 2026-07-20T17:34:54Z

+            if (SqlDbTypeExtensions.Json == effectiveType)
+            {
+                effectiveType = SqlDbType.NVarChar;
+            }


Treating Json as NVarChar might allow modification of individual characters via SqlDataRecord.SetChars. If that's true, it'd be possible to send invalid JSON strings to SQL Server.

Similar principle might also apply to VarBinary - would it allow arbitrary rewrites of the vector header via SetBytes?

edwardneal · 2026-07-20T17:37:03Z

+            using (SqlCommand createType = new($"CREATE TYPE {tvpTypeName} AS TABLE (Id int, Data json)", connection))
+            {
+                createType.ExecuteNonQuery();
+            }


We have an RAII primitive for this.

(here and elsewhere.)

edwardneal · 2026-07-20T17:40:50Z

It's also worth noting that TVPs can be sent using DataTables, and this uses a different pathway to detect the type.

edwardneal · 2026-07-20T17:52:21Z

                        result = GetSqlValue(getters, ordinal, metaData);
                        break;
+                    case SqlDbTypeExtensions.Json:
+                        result = new SqlString(GetString_Unchecked(getters, ordinal));


What would be the motivation for exposing this as a SqlString (or, further up, a string) rather than a SqlJson?

edwardneal · 2026-07-20T17:54:24Z

            );

+            // Vector values are transferred as their raw TDS payload bytes (8-byte header + elements).
+            if (SqlDbTypeExtensions.Vector == metaData.SqlDbType && value is ISqlVector vectorValue)


What happens if I pass null / SqlVector<float>.Null to this method? At a glance, I don't think it'll pass value is ISqlVector vectorValue and thus that it'll leave the previous value in place.

edwardneal · 2026-07-20T18:13:04Z

+                    // The SQL Server TVP protocol carries a JSON column as NVARCHAR(max) + collation,
+                    // with the value sent as UTF-16 (the server converts it to JSON on insert). This
+                    // matches the Microsoft JDBC driver's TVP/bulk representation of JSON columns.
+                    stateObj.WriteByte(TdsEnums.SQLNVARCHAR);


This is where I hit a wall a couple of days ago.

Encoding this as an SQLNVARCHAR will double the number of bytes needed to transmit ASCII text over the network. If we send SqlJson instances as a scalar parameter we encode the value to UTF8, so I'd tried to send it as a SQLJSON type (which is a PLP type.) This triggered some unusual behaviour within SQL Server though, are you able to replicate this?

edwardneal · 2026-07-20T18:15:14Z

+
+                SqlDataRecord row1 = new SqlDataRecord(metadata);
+                row1.SetInt32(0, 1);
+                row1.SetValue(1, SqlVector<float>.CreateNull(s_vectorDimensions)); // NULL vector


What happens if we call row1.SetDBNull?

edwardneal · 2026-07-20T18:19:45Z

+                };
+                SqlDataRecord row = new SqlDataRecord(metadata);
+                row.SetInt32(0, 1);
+                row.SetValue(1, new SqlVector<float>(values));


API nitpick: it'd be nice to have SqlDataRecord.SetVector<T> and SqlDataRecord.SetJson...

apoorvdeshmukh added 2 commits July 20, 2026 18:56

Add TVP support for json datatype

5bab824

TVP support for vectors

7fba6d8

Copilot AI review requested due to automatic review settings July 20, 2026 16:18

github-project-automation Bot added this to SqlClient Board Jul 20, 2026

github-project-automation Bot moved this to To triage in SqlClient Board Jul 20, 2026

Copilot started reviewing on behalf of apoorvdeshmukh July 20, 2026 16:18 View session

Copilot AI reviewed Jul 20, 2026

View reviewed changes

edwardneal reviewed Jul 20, 2026

View reviewed changes

		WriteUnsignedInt(_defaultCollation._info, stateObj);
		stateObj.WriteByte(_defaultCollation._sortId);

		// JSON and Vector are outside the contiguous SqlDbType range covered by this map;
		// JSON accepts the same setters as NVarChar (string) and Vector the same as VarBinary (bytes).

Uh oh!

Conversation

apoorvdeshmukh commented Jul 20, 2026

Description

Issues

Testing

Guidelines

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

codecov Bot commented Jul 20, 2026

Codecov Report

Uh oh!

edwardneal left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants