-
Notifications
You must be signed in to change notification settings - Fork 0
/
WikiCohereEnglishEmbeddingOnlyScenario.cs
409 lines (356 loc) · 20.2 KB
/
WikiCohereEnglishEmbeddingOnlyScenario.cs
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
using Microsoft.Azure.Cosmos;
using Microsoft.Extensions.Configuration;
using Newtonsoft.Json;
using System.Collections.Concurrent;
using System.Collections.ObjectModel;
namespace VectorIndexScenarioSuite
{
internal class WikiCohereEmbeddingOnlyDocument
{
// This should be the same as the PARTITION_KEY_PATH in the WikiCohereEnglishEmbeddingOnlyScenario.
[JsonProperty(PropertyName = "id")]
public string Id { get; }
// This should be the same as the EMBEDDING_PATH in the WikiCohereEnglishEmbeddingOnlyScenario.
[JsonProperty(PropertyName = "embedding")]
private float[] Embedding { get; }
public WikiCohereEmbeddingOnlyDocument(string id, float[] embedding)
{
this.Id = id;
this.Embedding = embedding;
}
}
internal class WikiCohereEnglishEmbeddingOnlyScenario : Scenario
{
private const string PARTITION_KEY_PATH = "/id";
private const string EMBEDDING_COLOUMN = "embedding";
private const string EMBEDDING_PATH = $"/{EMBEDDING_COLOUMN}";
private const VectorDataType EMBEDDING_DATA_TYPE = VectorDataType.Float32;
private const DistanceFunction EMBEDDING_DISTANCE_FUNCTION = DistanceFunction.Cosine;
private const int EMBEDDING_DIMENSIONS = 768;
private const int MAX_PHYSICAL_PARTITION_COUNT = 56;
private const string BASE_DATA_FILE = "wikipedia_base";
private const string QUERY_FILE = "wikipedia_query";
private const string GROUND_TRUTH_FILE = "wikipedia_truth";
private const string BINARY_FILE_EXT = "fbin";
private static readonly string RUN_NAME = "wiki-cohere-en-embeddingonly-" +
Guid.NewGuid();
/* Known Slices */
private const int HUNDRED_THOUSAND = 100000;
private const int ONE_MILLION = 1000000;
private const int THIRTY_FIVE_MILLION = 35000000;
/* Map 'K' -> Neighbor Results
* Neighbor Results:
* Map 'QueryId' -> List of neighbor IdWithSimilarityScore
*/
private ConcurrentDictionary<int, ConcurrentDictionary<string, List<IdWithSimilarityScore>>> queryRecallResults;
/* Map 'K' -> ScenarioMetrics (RU and Latency) */
private ConcurrentDictionary<int, ScenarioMetrics> queryMetrics;
private ScenarioMetrics ingestionMetrics;
public WikiCohereEnglishEmbeddingOnlyScenario(IConfiguration configurations) :
base(configurations, ComputeInitialAndFinalThroughput(configurations).Item1)
{
this.queryRecallResults = new ConcurrentDictionary<int, ConcurrentDictionary<string, List<IdWithSimilarityScore>>>();
this.queryMetrics = new ConcurrentDictionary<int, ScenarioMetrics>();
this.K_VALS = configurations.GetSection("AppSettings:scenario:kValues").Get<int[]>() ??
throw new ArgumentNullException("AppSettings:scenario:kValues");
this.ingestionMetrics = new ScenarioMetrics(0);
for(int kI = 0; kI < K_VALS.Length; kI++)
{
this.queryRecallResults.TryAdd(K_VALS[kI], new ConcurrentDictionary<string, List<IdWithSimilarityScore>>());
this.queryMetrics.TryAdd(K_VALS[kI], new ScenarioMetrics(0));
}
}
public override void Setup()
{
this.CosmosContainer.ReplaceThroughputAsync(ComputeInitialAndFinalThroughput(this.Configurations).Item2).Wait();
}
public override async Task Run()
{
/* WikiCohereEnglishScenario is a simple scenario with following steps :
* 1) Bulk Ingest 'scenario:slice' number of documents into Cosmos container.
* 2) Query Cosmos container for a query-set and calcualte recall for Nearest Neighbor search.
*/
bool runIngestion = Convert.ToBoolean(this.Configurations["AppSettings:scenario:runIngestion"]);
if(runIngestion)
{
await PerformIngestion();
}
bool runQuery = Convert.ToBoolean(this.Configurations["AppSettings:scenario:runQuery"]);
if(runQuery)
{
bool performWarmup = Convert.ToBoolean(this.Configurations["AppSettings:scenario:warmup:enabled"]);
if (performWarmup)
{
int numWarmupQueries = Convert.ToInt32(this.Configurations["AppSettings:scenario:warmup:numWarmupQueries"]);
Console.WriteLine($"Performing {numWarmupQueries} queries for Warmup.");
await PerformQuery(true /* isWarmup */, numWarmupQueries, 10 /*KVal*/, GetBaseDataPath());
}
int totalQueryVectors = BigANNBinaryFormat.GetBinaryDataHeader(GetQueryDataPath()).Item1;
for (int kI = 0; kI < K_VALS.Length; kI++)
{
Console.WriteLine($"Performing {totalQueryVectors} queries for Recall/RU/Latency stats for K: {K_VALS[kI]}.");
await PerformQuery(false /* isWarmup */, totalQueryVectors, K_VALS[kI] /*KVal*/, GetQueryDataPath());
}
}
}
public override void Stop()
{
bool runQuery = Convert.ToBoolean(this.Configurations["AppSettings:scenario:runQuery"]);
bool computeRecall = Convert.ToBoolean(this.Configurations["AppSettings:scenario:computeRecall"]);
if (runQuery && computeRecall)
{
Console.WriteLine("Computing Recall.");
GroundTruthValidator groundTruthValidator = new GroundTruthValidator(
GroundTruthFileType.Binary,
GetGroundTruthDataPath());
for (int kI = 0; kI < K_VALS.Length; kI++)
{
int kVal = K_VALS[kI];
float recall = groundTruthValidator.ComputeRecall(kVal, this.queryRecallResults[kVal]);
Console.WriteLine($"Recall for K = {kVal} is {recall}.");
}
}
bool computeLatencyAndRUStats = Convert.ToBoolean(this.Configurations["AppSettings:scenario:computeLatencyAndRUStats"]);
if (computeLatencyAndRUStats)
{
bool runIngestion = Convert.ToBoolean(this.Configurations["AppSettings:scenario:runIngestion"]);
ComputeLatencyAndRUStats(runIngestion, runQuery);
}
}
private static (int, int) ComputeInitialAndFinalThroughput(IConfiguration configurations)
{
// For wiki-coherscenario, we are starting with :
// 1) For upto 1M embedding, Collection Create throughput of 400 RU, bumped to 10,000 RU.
// 2) For 35M embedding, Collection Create throughput of 40,000 RU, bumped to 70,000 RU.
// This is because we want 1 physical partition in scenrio 1 and 7 physical partitions in scenario 2 (to reduce query fanout).
int sliceCount = Convert.ToInt32(configurations["AppSettings:scenario:sliceCount"]);
switch (sliceCount)
{
case HUNDRED_THOUSAND:
case ONE_MILLION:
return (400, 10000);
case THIRTY_FIVE_MILLION:
return (40000, 70000);
default:
throw new ArgumentException("Invalid slice count.");
}
}
private void ComputeLatencyAndRUStats(bool runIngestion, bool runQuery)
{
if(runIngestion)
{
Console.WriteLine($"Ingestion Metrics:");
Console.WriteLine($"RU Consumption: {this.ingestionMetrics.GetRequestUnitStatistics()}");
}
if (runQuery)
{
Console.WriteLine($"Query Metrics:");
for(int kI = 0; kI < K_VALS.Length; kI++)
{
int kVal = K_VALS[kI];
Console.WriteLine($"K = {kVal}");
ScenarioMetrics metrics = this.queryMetrics[kVal];
Console.WriteLine($"RU Consumption: {metrics.GetRequestUnitStatistics()}");
Console.WriteLine($"Client Latency Stats in Milliseconds: [ {metrics.GetClientLatencyStatistics()} ]");
Console.WriteLine($"Server Latency Stats in Milliseconds: [ {metrics.GetServerLatencyStatistics()} ]");
}
}
}
private async Task PerformIngestion()
{
int numBulkIngestionBatchCount = Convert.ToInt32(this.Configurations["AppSettings:scenario:numBulkIngestionBatchCount"]);
int totalVectors = Convert.ToInt32(this.Configurations["AppSettings:scenario:sliceCount"]);
if (totalVectors % numBulkIngestionBatchCount != 0)
{
throw new ArgumentException("Total vectors should be evenly divisible by numBulkIngestionBatchCount");
}
int numVectorsPerRange = totalVectors / numBulkIngestionBatchCount;
this.ingestionMetrics = new ScenarioMetrics(totalVectors);
var tasks = new List<Task>(numBulkIngestionBatchCount);
for (int rangeIndex = 0; rangeIndex < numBulkIngestionBatchCount; rangeIndex++)
{
int startVectorId = rangeIndex * numVectorsPerRange ;
Console.WriteLine(
$"Starting ingestion for range: {rangeIndex} with start vectorId: [{startVectorId}, " +
$"{startVectorId + numVectorsPerRange})");
tasks.Add(BulkIngestDataForRange(startVectorId, numVectorsPerRange));
}
await Task.WhenAll(tasks);
}
private async Task PerformQuery(bool isWarmup, int numQueries, int KVal, string dataPath)
{
if(!isWarmup)
{
this.queryMetrics[KVal] = new ScenarioMetrics(numQueries);
}
await foreach ((int vectorId, float[] vector) in
BigANNBinaryFormat.GetBinaryDataAsync(dataPath, BinaryDataType.Float32, 0 /* startVectorId */, numQueries))
{
var queryDefinition = ConstructQueryDefinition(KVal, vector);
bool retryQueryOnFailureForLatencyMeasurement;
do
{
FeedIterator<IdWithSimilarityScore> queryResultSetIterator =
this.CosmosContainer.GetItemQueryIterator<IdWithSimilarityScore>(queryDefinition,
// Issue parallel queries to all partitions, capping this to MAX_PHYSICAL_PARTITION_COUNT but can be tuned based on change in setup.
requestOptions: new QueryRequestOptions { MaxConcurrency = (MAX_PHYSICAL_PARTITION_COUNT) });
retryQueryOnFailureForLatencyMeasurement = false;
while (queryResultSetIterator.HasMoreResults)
{
var queryResponse = await queryResultSetIterator.ReadNextAsync();
if (!isWarmup)
{
// If we are computing latency and RU stats, don't consider any query with failed requests (implies it was throttled).
bool computeLatencyAndRUStats = Convert.ToBoolean(this.Configurations["AppSettings:scenario:computeLatencyAndRUStats"]);
if (computeLatencyAndRUStats && queryResponse.Diagnostics.GetFailedRequestCount() > 0)
{
Console.WriteLine($"Retrying for vectorId : {vectorId}.");
retryQueryOnFailureForLatencyMeasurement = true;
break;
}
if (!retryQueryOnFailureForLatencyMeasurement && queryResponse.Count > 0)
{
// Get Client time before doing any more work. Validated this matches stopwatch time.
// The second iteration does not have meaningful RU and Latency numbers.
if (queryResponse.RequestCharge > 0)
{
this.queryMetrics[KVal].AddRequestUnitMeasurement(
queryResponse.RequestCharge);
this.queryMetrics[KVal].AddClientLatencyMeasurement(
queryResponse.Diagnostics.GetClientElapsedTime().TotalMilliseconds);
}
if (!this.queryRecallResults[KVal].ContainsKey(vectorId.ToString()))
{
this.queryRecallResults[KVal].TryAdd(vectorId.ToString(), new List<IdWithSimilarityScore>(KVal));
}
var results = this.queryRecallResults[KVal][vectorId.ToString()];
foreach (var idWithScoreResponse in queryResponse)
{
results.Add(idWithScoreResponse);
}
// Similarly, QueryMetrics is null for second and subsequent pages of query results.
if (queryResponse.Diagnostics.GetQueryMetrics() != null)
{
this.queryMetrics[KVal].AddServerLatencyMeasurement(
queryResponse.Diagnostics.GetQueryMetrics().CumulativeMetrics.TotalTime.TotalMilliseconds);
}
}
}
}
int vectorCount = vectorId + 1;
if (vectorCount % COSMOSDB_MAX_BATCH_SIZE == 0)
{
double percentage = ((double)vectorId / numQueries) * 100;
Console.WriteLine($"Finished querying {percentage.ToString("F2")}% ");
}
}
while ( retryQueryOnFailureForLatencyMeasurement );
}
}
private QueryDefinition ConstructQueryDefinition(int K, float[] queryVector)
{
string queryText = $"SELECT TOP {K} c.id, VectorDistance(c.{EMBEDDING_COLOUMN}, @vectorEmbedding) AS similarityScore " +
$"FROM c ORDER BY VectorDistance(c.{EMBEDDING_COLOUMN}, @vectorEmbedding, false)";
;
return new QueryDefinition(queryText).WithParameter("@vectorEmbedding", queryVector);
}
private async Task BulkIngestDataForRange(int startVectorId, int numVectorsToIngest)
{
// The batches that the SDK creates to optimize throughput have a current maximum of 2Mb or 100 operations per batch,
List<Task> ingestTasks = new List<Task>(COSMOSDB_MAX_BATCH_SIZE);
string errorLogBasePath = this.Configurations["AppSettings:errorLogBasePath"] ??
throw new ArgumentNullException("AppSettings:errorLogBasePath");
string logFilePath = Path.Combine(errorLogBasePath, $"{RUN_NAME}-ingest.log");
int totalVectorsIngested = 0;
await foreach ((int vectorId, float[] vector) in BigANNBinaryFormat.GetBinaryDataAsync(GetBaseDataPath(), BinaryDataType.Float32, startVectorId, numVectorsToIngest))
{
var createTask = this.CosmosContainerWithBulkClient.CreateItemAsync<WikiCohereEmbeddingOnlyDocument>(
new WikiCohereEmbeddingOnlyDocument(vectorId.ToString(), vector), new PartitionKey(vectorId.ToString())).ContinueWith(async itemResponse =>
{
if (!itemResponse.IsCompletedSuccessfully)
{
Console.WriteLine($"Insert failed for id: {vectorId}.");
// Log the error to a file
string errorLogMessage = $"Error ingesting vectorId: {vectorId}, " +
$"Error: {itemResponse.Exception.InnerException.Message}";
await LogErrorToFile(logFilePath, errorLogMessage);
}
else
{
// Given we are doing bulk ingestion which is optimized for throughput and not latency, we are not mesuring latency numbers.
this.ingestionMetrics.AddRequestUnitMeasurement(itemResponse.Result.RequestCharge);
}
}).Unwrap();
ingestTasks.Add(createTask);
if (ingestTasks.Count == COSMOSDB_MAX_BATCH_SIZE)
{
await Task.WhenAll(ingestTasks);
ingestTasks.Clear();
totalVectorsIngested += COSMOSDB_MAX_BATCH_SIZE;
double percentage = ((double)totalVectorsIngested / numVectorsToIngest) * 100;
Console.WriteLine($"Finished ingesting {percentage.ToString("F2")}% " +
$"for Range [{startVectorId},{startVectorId + numVectorsToIngest}).");
}
}
if (ingestTasks.Count > 0)
{
await Task.WhenAll(ingestTasks);
totalVectorsIngested += ingestTasks.Count;
ingestTasks.Clear();
Console.WriteLine($"Ingested {totalVectorsIngested} documents for range with start vectorId {startVectorId}");
}
}
private string GetQueryDataPath()
{
string directory = this.Configurations["AppSettings:dataFilesBasePath"] ??
throw new ArgumentNullException("AppSettings:dataFilesBasePath");
string fileName = $"{QUERY_FILE}.{BINARY_FILE_EXT}";
return Path.Combine(directory, fileName);
}
private string GetBaseDataPath()
{
string directory = this.Configurations["AppSettings:dataFilesBasePath"] ??
throw new ArgumentNullException("AppSettings:dataFilesBasePath");
string fileName = $"{BASE_DATA_FILE}_{this.Configurations["AppSettings:scenario:sliceCount"]}.{BINARY_FILE_EXT}";
return Path.Combine(directory, fileName);
}
private string GetGroundTruthDataPath()
{
string directory = this.Configurations["AppSettings:dataFilesBasePath"] ??
throw new ArgumentNullException("AppSettings:dataFilesBasePath");
string fileName = $"{GROUND_TRUTH_FILE}_{this.Configurations["AppSettings:scenario:sliceCount"]}";
return Path.Combine(directory, fileName);
}
protected override ContainerProperties GetContainerSpec(string containerName)
{
ContainerProperties properties = new ContainerProperties(id: containerName, partitionKeyPath: PARTITION_KEY_PATH)
{
VectorEmbeddingPolicy = new VectorEmbeddingPolicy(new Collection<Embedding>(new List<Embedding>()
{
new Embedding()
{
Path = EMBEDDING_PATH,
DataType = VectorDataType.Float32,
DistanceFunction = DistanceFunction.DotProduct,
Dimensions = 768,
}
})),
IndexingPolicy = new IndexingPolicy()
{
VectorIndexes = new()
{
new VectorIndexPath()
{
Path = EMBEDDING_PATH,
Type = VectorIndexType.DiskANN,
}
}
}
};
properties.IndexingPolicy.IncludedPaths.Add(new IncludedPath{ Path = "/" });
// Add EMBEDDING_PATH to excluded paths for scalar indexing.
properties.IndexingPolicy.ExcludedPaths.Add(new ExcludedPath { Path = EMBEDDING_PATH + "/*" });
return properties;
}
}
}