Add hardware usage to grpc API (#5333)

* Add hardware usage to grpc API * generate grpc docs
qdrant · Nov 5, 2024 · e35ec15 · e35ec15
1 parent 978c4bd
commit e35ec15
Show file tree

Hide file tree

Showing 9 changed files with 225 additions and 43 deletions.
diff --git a/docs/grpc/docs.md b/docs/grpc/docs.md
@@ -156,6 +156,7 @@
     - [GetResponse](#qdrant-GetResponse)
     - [GroupId](#qdrant-GroupId)
     - [GroupsResult](#qdrant-GroupsResult)
+    - [HardwareUsage](#qdrant-HardwareUsage)
     - [HasIdCondition](#qdrant-HasIdCondition)
     - [HasVectorCondition](#qdrant-HasVectorCondition)
     - [Image](#qdrant-Image)
@@ -2218,6 +2219,7 @@ The JSON representation for `Value` is a JSON value.
 | ----- | ---- | ----- | ----------- |
 | result | [CountResult](#qdrant-CountResult) |  |  |
 | time | [double](#double) |  | Time spent to process |
+| usage | [HardwareUsage](#qdrant-HardwareUsage) | optional |  |
 
 
 
@@ -2397,6 +2399,7 @@ The JSON representation for `Value` is a JSON value.
 | ----- | ---- | ----- | ----------- |
 | result | [BatchResult](#qdrant-BatchResult) | repeated |  |
 | time | [double](#double) |  | Time spent to process |
+| usage | [HardwareUsage](#qdrant-HardwareUsage) | optional |  |
 
 
 
@@ -2457,6 +2460,7 @@ The JSON representation for `Value` is a JSON value.
 | ----- | ---- | ----- | ----------- |
 | result | [ScoredPoint](#qdrant-ScoredPoint) | repeated |  |
 | time | [double](#double) |  | Time spent to process |
+| usage | [HardwareUsage](#qdrant-HardwareUsage) | optional |  |
 
 
 
@@ -2756,6 +2760,21 @@ Additionally, the first and last points of each GeoLineString must be the same.
 
 
 
+<a name="qdrant-HardwareUsage"></a>
+
+### HardwareUsage
+
+
+
+| Field | Type | Label | Description |
+| ----- | ---- | ----- | ----------- |
+| cpu | [uint64](#uint64) |  |  |
+
+
+
+
+
+
 <a name="qdrant-HasIdCondition"></a>
 
 ### HasIdCondition
@@ -3504,6 +3523,7 @@ For example, if `oversampling` is 2.4 and `limit` is 100, then 240 vectors will
 | ----- | ---- | ----- | ----------- |
 | result | [BatchResult](#qdrant-BatchResult) | repeated |  |
 | time | [double](#double) |  | Time spent to process |
+| usage | [HardwareUsage](#qdrant-HardwareUsage) | optional |  |
 
 
 
@@ -3520,6 +3540,7 @@ For example, if `oversampling` is 2.4 and `limit` is 100, then 240 vectors will
 | ----- | ---- | ----- | ----------- |
 | result | [GroupsResult](#qdrant-GroupsResult) |  |  |
 | time | [double](#double) |  | Time spent to process |
+| usage | [HardwareUsage](#qdrant-HardwareUsage) | optional |  |
 
 
 
@@ -3596,6 +3617,7 @@ For example, if `oversampling` is 2.4 and `limit` is 100, then 240 vectors will
 | ----- | ---- | ----- | ----------- |
 | result | [ScoredPoint](#qdrant-ScoredPoint) | repeated |  |
 | time | [double](#double) |  | Time spent to process |
+| usage | [HardwareUsage](#qdrant-HardwareUsage) | optional |  |
 
 
 
@@ -3664,6 +3686,7 @@ For example, if `oversampling` is 2.4 and `limit` is 100, then 240 vectors will
 | ----- | ---- | ----- | ----------- |
 | result | [BatchResult](#qdrant-BatchResult) | repeated |  |
 | time | [double](#double) |  | Time spent to process |
+| usage | [HardwareUsage](#qdrant-HardwareUsage) | optional |  |
 
 
 
@@ -3680,6 +3703,7 @@ For example, if `oversampling` is 2.4 and `limit` is 100, then 240 vectors will
 | ----- | ---- | ----- | ----------- |
 | result | [GroupsResult](#qdrant-GroupsResult) |  |  |
 | time | [double](#double) |  | Time spent to process |
+| usage | [HardwareUsage](#qdrant-HardwareUsage) | optional |  |
 
 
 
@@ -3779,6 +3803,7 @@ For example, if `oversampling` is 2.4 and `limit` is 100, then 240 vectors will
 | ----- | ---- | ----- | ----------- |
 | result | [ScoredPoint](#qdrant-ScoredPoint) | repeated |  |
 | time | [double](#double) |  | Time spent to process |
+| usage | [HardwareUsage](#qdrant-HardwareUsage) | optional |  |
 
 
 
@@ -3956,6 +3981,7 @@ For example, if `oversampling` is 2.4 and `limit` is 100, then 240 vectors will
 | ----- | ---- | ----- | ----------- |
 | result | [BatchResult](#qdrant-BatchResult) | repeated |  |
 | time | [double](#double) |  | Time spent to process |
+| usage | [HardwareUsage](#qdrant-HardwareUsage) | optional |  |
 
 
 
@@ -3972,6 +3998,7 @@ For example, if `oversampling` is 2.4 and `limit` is 100, then 240 vectors will
 | ----- | ---- | ----- | ----------- |
 | result | [GroupsResult](#qdrant-GroupsResult) |  |  |
 | time | [double](#double) |  | Time spent to process |
+| usage | [HardwareUsage](#qdrant-HardwareUsage) | optional |  |
 
 
 
@@ -4006,6 +4033,7 @@ For example, if `oversampling` is 2.4 and `limit` is 100, then 240 vectors will
 | ----- | ---- | ----- | ----------- |
 | result | [SearchMatrixOffsets](#qdrant-SearchMatrixOffsets) |  |  |
 | time | [double](#double) |  | Time spent to process |
+| usage | [HardwareUsage](#qdrant-HardwareUsage) | optional |  |
 
 
 
@@ -4054,6 +4082,7 @@ For example, if `oversampling` is 2.4 and `limit` is 100, then 240 vectors will
 | ----- | ---- | ----- | ----------- |
 | result | [SearchMatrixPairs](#qdrant-SearchMatrixPairs) |  |  |
 | time | [double](#double) |  | Time spent to process |
+| usage | [HardwareUsage](#qdrant-HardwareUsage) | optional |  |
 
 
 
@@ -4168,6 +4197,7 @@ For example, if `oversampling` is 2.4 and `limit` is 100, then 240 vectors will
 | ----- | ---- | ----- | ----------- |
 | result | [ScoredPoint](#qdrant-ScoredPoint) | repeated |  |
 | time | [double](#double) |  | Time spent to process |
+| usage | [HardwareUsage](#qdrant-HardwareUsage) | optional |  |
 
 
 

diff --git a/lib/api/src/grpc/conversions.rs b/lib/api/src/grpc/conversions.rs
@@ -3,6 +3,7 @@ use std::str::FromStr as _;
 use std::time::Instant;
 
 use chrono::{NaiveDateTime, Timelike};
+use common::counter::hardware_accumulator::HwMeasurementAcc;
 use itertools::Itertools;
 use segment::common::operation_error::OperationError;
 use segment::data_types::index::{
@@ -21,7 +22,7 @@ use super::qdrant::{
     raw_query, start_from, BinaryQuantization, BoolIndexParams, CompressionRatio,
     DatetimeIndexParams, DatetimeRange, Direction, FacetHit, FacetHitInternal, FacetValue,
     FacetValueInternal, FieldType, FloatIndexParams, GeoIndexParams, GeoLineString, GroupId,
-    HasVectorCondition, KeywordIndexParams, LookupLocation, MultiVectorComparator,
+    HardwareUsage, HasVectorCondition, KeywordIndexParams, LookupLocation, MultiVectorComparator,
     MultiVectorConfig, OrderBy, OrderValue, Range, RawVector, RecommendStrategy, RetrievedPoint,
     SearchMatrixPair, SearchPointGroups, SearchPoints, ShardKeySelector, SparseIndices, StartFrom,
     UuidIndexParams, VectorsOutput, WithLookup,
@@ -2129,3 +2130,11 @@ impl From<rest::SearchMatrixPair> for SearchMatrixPair {
         }
     }
 }
+
+impl From<HwMeasurementAcc> for HardwareUsage {
+    fn from(value: HwMeasurementAcc) -> Self {
+        Self {
+            cpu: value.get_cpu() as u64,
+        }
+    }
+}
diff --git a/lib/api/src/grpc/proto/points.proto b/lib/api/src/grpc/proto/points.proto
@@ -817,21 +817,25 @@ message GroupsResult {
 message SearchResponse {
   repeated ScoredPoint result = 1;
   double time = 2; // Time spent to process
+  optional HardwareUsage usage = 3;
 }
 
 message QueryResponse {
   repeated ScoredPoint result = 1;
   double time = 2; // Time spent to process
+  optional HardwareUsage usage = 3;
 }
 
 message QueryBatchResponse {
   repeated BatchResult result = 1;
   double time = 2; // Time spent to process
+  optional HardwareUsage usage = 3;
 }
 
 message QueryGroupsResponse {
   GroupsResult result = 1;
   double time = 2; // Time spent to process
+  optional HardwareUsage usage = 3;
 }
 
 message BatchResult {
@@ -841,16 +845,19 @@ message BatchResult {
 message SearchBatchResponse {
   repeated BatchResult result = 1;
   double time = 2; // Time spent to process
+  optional HardwareUsage usage = 3;
 }
 
 message SearchGroupsResponse {
   GroupsResult result = 1;
   double time = 2; // Time spent to process
+  optional HardwareUsage usage = 3;
 }
 
 message CountResponse {
   CountResult result = 1;
   double time = 2; // Time spent to process
+  optional HardwareUsage usage = 3;
 }
 
 message ScrollResponse {
@@ -880,26 +887,31 @@ message GetResponse {
 message RecommendResponse {
   repeated ScoredPoint result = 1;
   double time = 2; // Time spent to process
+  optional HardwareUsage usage = 3;
 }
 
 message RecommendBatchResponse {
   repeated BatchResult result = 1;
   double time = 2; // Time spent to process
+  optional HardwareUsage usage = 3;
 }
 
 message DiscoverResponse {
   repeated ScoredPoint result = 1;
   double time = 2; // Time spent to process
+  optional HardwareUsage usage = 3;
 }
 
 message DiscoverBatchResponse {
   repeated BatchResult result = 1;
   double time = 2; // Time spent to process
+  optional HardwareUsage usage = 3;
 }
 
 message RecommendGroupsResponse {
   GroupsResult result = 1;
   double time = 2; // Time spent to process
+  optional HardwareUsage usage = 3;
 }
 
 message UpdateBatchResponse {
@@ -915,11 +927,13 @@ message FacetResponse {
 message SearchMatrixPairsResponse {
   SearchMatrixPairs result = 1;
   double time = 2; // Time spent to process
+  optional HardwareUsage usage = 3;
 }
 
 message SearchMatrixOffsetsResponse {
   SearchMatrixOffsets result = 1;
   double time = 2; // Time spent to process
+  optional HardwareUsage usage = 3;
 }
 
 // ---------------------------------------------
@@ -1077,3 +1091,11 @@ message GeoPoint {
   double lon = 1;
   double lat = 2;
 }
+
+// ---------------------------------------------
+// ------------ Hardware measurements ----------
+// ---------------------------------------------
+
+message HardwareUsage {
+  uint64 cpu = 1;
+}
diff --git a/lib/api/src/grpc/proto/points_internal_service.proto b/lib/api/src/grpc/proto/points_internal_service.proto
@@ -287,6 +287,7 @@ message QueryResultInternal {
 message QueryBatchResponseInternal {
   repeated QueryResultInternal results = 1;
   double time = 2; // Time spent to process
+  optional HardwareUsage usage = 5;
 }
 
 message FacetCountsInternal {