private void ApplyPrecomputedBatchForNewIndex(Index index, AbstractViewGenerator generator, int pageSize, CancellationTokenSource cts)
{
PrecomputedIndexingBatch result = null;
var docsToIndex = new List <JsonDocument>();
TransactionalStorage.Batch(actions =>
{
var query = GetQueryForAllMatchingDocumentsForIndex(generator);
using (DocumentCacher.SkipSetDocumentsInDocumentCache())
using (var linked = CancellationTokenSource.CreateLinkedTokenSource(cts.Token, WorkContext.CancellationToken))
using (var op = new QueryActions.DatabaseQueryOperation(Database, Constants.DocumentsByEntityNameIndex, new IndexQuery
{
Query = query,
PageSize = pageSize
}, actions, linked)
{
ShouldSkipDuplicateChecking = true
})
{
op.Init();
//if we are working on a test index, apply the optimization anyway, as the index is capped by small number of results
if (index.IsTestIndex == false && op.Header.TotalResults > pageSize)
{
// we don't apply this optimization if the total number of results
// to index is more than the max numbers to index in a single batch.
// The idea here is that we need to keep the amount
// of memory we use to a manageable level even when introducing a new index to a BIG
// database
try
{
cts.Cancel();
// we have to run just a little bit of the query to properly setup the disposal
op.Execute(o => { });
}
catch (OperationCanceledException)
{
}
return;
}
if (Log.IsDebugEnabled)
{
Log.Debug("For new index {0}, using precomputed indexing batch optimization for {1} docs",
index, op.Header.TotalResults);
}
var totalLoadedDocumentSize = 0;
const int totalSizeToCheck = 16 * 1024 * 1024; //16MB
var localLoadedDocumentSize = 0;
op.Execute(document =>
{
var metadata = document.Value <RavenJObject>(Constants.Metadata);
var key = metadata.Value <string>("@id");
var etag = Etag.Parse(metadata.Value <string>("@etag"));
var lastModified = DateTime.Parse(metadata.Value <string>(Constants.LastModified));
document.Remove(Constants.Metadata);
var serializedSizeOnDisk = metadata.Value <int>(Constants.SerializedSizeOnDisk);
metadata.Remove(Constants.SerializedSizeOnDisk);
var doc = new JsonDocument
{
DataAsJson = document,
Etag = etag,
Key = key,
SerializedSizeOnDisk = serializedSizeOnDisk,
LastModified = lastModified,
SkipDeleteFromIndex = true,
Metadata = metadata
};
docsToIndex.Add(doc);
totalLoadedDocumentSize += serializedSizeOnDisk;
localLoadedDocumentSize += serializedSizeOnDisk;
if (totalLoadedDocumentSize > Database.Configuration.MaxPrecomputedBatchTotalDocumentSizeInBytes)
{
var error = string.Format(
@"Aborting applying precomputed batch for index id: {0}, name: {1}
because we have {2}mb of documents that were fetched
and the configured max data to fetch is {3}mb",
index.indexId, index.PublicName, totalLoadedDocumentSize,
Database.Configuration.MaxPrecomputedBatchTotalDocumentSizeInBytes / 1024 / 1024);
//we are aborting operation, so don't keep the references
docsToIndex.Clear();
throw new TotalDataSizeExceededException(error);
}
if (localLoadedDocumentSize <= totalSizeToCheck)
{
return;
}
localLoadedDocumentSize = 0;
if (Database.Configuration.MemoryLimitForProcessingInMb > MemoryStatistics.AvailableMemoryInMb)
{
var error = string.Format(
@"Aborting applying precomputed batch for index id: {0}, name: {1}
because we have {2}mb of available memory and the available memory for processing is: {3}mb",
index.indexId, index.PublicName,
MemoryStatistics.AvailableMemoryInMb, Database.Configuration.MemoryLimitForProcessingInMb);
//we are aborting operation, so don't keep the references
docsToIndex.Clear();
throw new TotalDataSizeExceededException(error);
}
});
result = new PrecomputedIndexingBatch
{
LastIndexed = op.Header.IndexEtag,
LastModified = op.Header.IndexTimestamp,
Documents = docsToIndex,
Index = index
};
}
});
if (result != null && result.Documents != null && result.Documents.Count >= 0)
{
using (var linked = CancellationTokenSource.CreateLinkedTokenSource(cts.Token, WorkContext.CancellationToken))
{
Database.IndexingExecuter.IndexPrecomputedBatch(result, linked.Token);
if (index.IsTestIndex)
{
TransactionalStorage.Batch(accessor => accessor.Indexing.TouchIndexEtag(index.IndexId));
}
}
}
}
private ReducingPerformanceStats MultiStepReduce(IndexToWorkOn index, List <string> keysToReduce,
AbstractViewGenerator viewGenerator, ConcurrentSet <object> itemsToDelete, bool skipIncreasingBatchSize, CancellationToken token)
{
var needToMoveToMultiStep = new HashSet <string>();
var alreadyMultiStep = new HashSet <string>();
transactionalStorage.Batch(actions =>
{
foreach (var localReduceKey in keysToReduce)
{
token.ThrowIfCancellationRequested();
var lastPerformedReduceType = actions.MapReduce.GetLastPerformedReduceType(index.IndexId, localReduceKey);
if (lastPerformedReduceType != ReduceType.MultiStep)
{
needToMoveToMultiStep.Add(localReduceKey);
}
if (lastPerformedReduceType == ReduceType.MultiStep)
{
alreadyMultiStep.Add(localReduceKey);
}
if (lastPerformedReduceType != ReduceType.SingleStep)
{
continue;
}
// we exceeded the limit of items to reduce in single step
// now we need to schedule reductions at level 0 for all map results with given reduce key
var mappedItems = actions.MapReduce.GetMappedBuckets(index.IndexId, localReduceKey, token).ToList();
foreach (var result in mappedItems.Select(x => new ReduceKeyAndBucket(x, localReduceKey)))
{
actions.MapReduce.ScheduleReductions(index.IndexId, 0, result);
}
}
});
var reducePerformance = new ReducingPerformanceStats(ReduceType.MultiStep);
for (int i = 0; i < 3; i++)
{
var level = i;
var reduceLevelStats = new ReduceLevelPeformanceStats()
{
Level = level,
Started = SystemTime.UtcNow,
};
var reduceParams = new GetItemsToReduceParams(
index.IndexId,
new HashSet <string>(keysToReduce),
level,
true,
itemsToDelete);
var gettingItemsToReduceDuration = new Stopwatch();
var scheduleReductionsDuration = new Stopwatch();
var removeReduceResultsDuration = new Stopwatch();
var storageCommitDuration = new Stopwatch();
bool retry = true;
while (retry && reduceParams.ReduceKeys.Count > 0)
{
var reduceBatchAutoThrottlerId = Guid.NewGuid();
try
{
transactionalStorage.Batch(actions =>
{
token.ThrowIfCancellationRequested();
actions.BeforeStorageCommit += storageCommitDuration.Start;
actions.AfterStorageCommit += storageCommitDuration.Stop;
var batchTimeWatcher = Stopwatch.StartNew();
reduceParams.Take = context.CurrentNumberOfItemsToReduceInSingleBatch;
int size = 0;
IList <MappedResultInfo> persistedResults;
var reduceKeys = new HashSet <string>(StringComparer.InvariantCultureIgnoreCase);
using (StopwatchScope.For(gettingItemsToReduceDuration))
{
persistedResults = actions.MapReduce.GetItemsToReduce(reduceParams, token);
foreach (var item in persistedResults)
{
reduceKeys.Add(item.ReduceKey);
size += item.Size;
}
}
if (persistedResults.Count == 0)
{
retry = false;
return;
}
var count = persistedResults.Count;
autoTuner.CurrentlyUsedBatchSizesInBytes.GetOrAdd(reduceBatchAutoThrottlerId, size);
if (Log.IsDebugEnabled)
{
if (persistedResults.Count > 0)
{
if (Log.IsDebugEnabled)
{
Log.Debug(() => string.Format("Found {0} results for keys [{1}] for index {2} at level {3} in {4}",
persistedResults.Count,
string.Join(", ", persistedResults.Select(x => x.ReduceKey).Distinct()),
index.Index.PublicName, level, batchTimeWatcher.Elapsed));
}
}
else
{
if (Log.IsDebugEnabled)
{
Log.Debug("No reduce keys found for {0}", index.Index.PublicName);
}
}
}
token.ThrowIfCancellationRequested();
var requiredReduceNextTimeSet = new HashSet <ReduceKeyAndBucket>(persistedResults.Select(x => new ReduceKeyAndBucket(x.Bucket, x.ReduceKey)), ReduceKeyAndBucketEqualityComparer.Instance);
using (StopwatchScope.For(removeReduceResultsDuration))
{
foreach (var mappedResultInfo in requiredReduceNextTimeSet)
{
token.ThrowIfCancellationRequested();
actions.MapReduce.RemoveReduceResults(index.IndexId, level + 1, mappedResultInfo.ReduceKey, mappedResultInfo.Bucket);
}
}
if (level != 2)
{
var reduceKeysAndBucketsSet = new HashSet <ReduceKeyAndBucket>(requiredReduceNextTimeSet.Select(x => new ReduceKeyAndBucket(x.Bucket / 1024, x.ReduceKey)), ReduceKeyAndBucketEqualityComparer.Instance);
using (StopwatchScope.For(scheduleReductionsDuration))
{
foreach (var reduceKeysAndBucket in reduceKeysAndBucketsSet)
{
token.ThrowIfCancellationRequested();
actions.MapReduce.ScheduleReductions(index.IndexId, level + 1, reduceKeysAndBucket);
}
}
}
token.ThrowIfCancellationRequested();
var reduceTimeWatcher = Stopwatch.StartNew();
var results = persistedResults.Where(x => x.Data != null)
.GroupBy(x => x.Bucket, x => JsonToExpando.Convert(x.Data))
.ToList();
var performance = context.IndexStorage.Reduce(index.IndexId, viewGenerator, results, level, context, actions, reduceKeys, persistedResults.Count);
context.MetricsCounters.ReducedPerSecond.Mark(results.Count());
reduceLevelStats.Add(performance);
var batchDuration = batchTimeWatcher.Elapsed;
if (Log.IsDebugEnabled)
{
Log.Debug("Indexed {0} reduce keys in {1} with {2} results for index {3} in {4} on level {5}", reduceKeys.Count, batchDuration, performance.ItemsCount, index.Index.PublicName, reduceTimeWatcher.Elapsed, level);
}
autoTuner.AutoThrottleBatchSize(count, size, batchDuration, skipIncreasingBatchSize);
});
}
finally
{
long _;
autoTuner.CurrentlyUsedBatchSizesInBytes.TryRemove(reduceBatchAutoThrottlerId, out _);
}
}
reduceLevelStats.Completed = SystemTime.UtcNow;
reduceLevelStats.Duration = reduceLevelStats.Completed - reduceLevelStats.Started;
reduceLevelStats.Operations.Add(PerformanceStats.From(IndexingOperation.Reduce_GetItemsToReduce, gettingItemsToReduceDuration.ElapsedMilliseconds));
reduceLevelStats.Operations.Add(PerformanceStats.From(IndexingOperation.Reduce_ScheduleReductions, scheduleReductionsDuration.ElapsedMilliseconds));
reduceLevelStats.Operations.Add(PerformanceStats.From(IndexingOperation.Reduce_RemoveReduceResults, removeReduceResultsDuration.ElapsedMilliseconds));
reduceLevelStats.Operations.Add(PerformanceStats.From(IndexingOperation.StorageCommit, storageCommitDuration.ElapsedMilliseconds));
reducePerformance.LevelStats.Add(reduceLevelStats);
}
// update new preformed multi step
UpdatePerformedReduceType(index.IndexId, needToMoveToMultiStep, ReduceType.MultiStep);
// already multi step,
// if the multi step keys were already removed,
// the reduce types for those keys also need to be removed
UpdatePerformedReduceType(index.IndexId, alreadyMultiStep, ReduceType.MultiStep, skipAdd: true);
return(reducePerformance);
}
private ReducingPerformanceStats SingleStepReduce(IndexToWorkOn index, List <string> keysToReduce, AbstractViewGenerator viewGenerator,
ConcurrentSet <object> itemsToDelete, bool skipIncreasingBatchSize, CancellationToken token)
{
var needToMoveToSingleStepQueue = new ConcurrentQueue <HashSet <string> >();
var alreadySingleStepQueue = new ConcurrentQueue <HashSet <string> >();
if (Log.IsDebugEnabled)
{
Log.Debug(() => $"Executing single step reducing for {keysToReduce.Count} keys [{string.Join(", ", keysToReduce)}]");
}
var batchTimeWatcher = Stopwatch.StartNew();
var reducingBatchThrottlerId = Guid.NewGuid();
var reducePerformanceStats = new ReducingPerformanceStats(ReduceType.SingleStep);
var reduceLevelStats = new ReduceLevelPeformanceStats
{
Started = SystemTime.UtcNow,
Level = 2
};
try
{
var parallelOperations = new ConcurrentQueue <ParallelBatchStats>();
var parallelProcessingStart = SystemTime.UtcNow;
if (context.Database.ThreadPool == null || context.RunReducing == false)
{
throw new OperationCanceledException();
}
context.Database.ThreadPool.ExecuteBatch(keysToReduce, enumerator =>
{
var parallelStats = new ParallelBatchStats
{
StartDelay = (long)(SystemTime.UtcNow - parallelProcessingStart).TotalMilliseconds
};
while (enumerator.MoveNext())
{
var localKeys = new HashSet <string>();
for (var i = 0; i < RavenThreadPool.DefaultPageSize; i++)
{
token.ThrowIfCancellationRequested();
localKeys.Add(enumerator.Current);
if (enumerator.MoveNext() == false)
{
break;
}
}
if (localKeys.Count == 0)
{
return;
}
var localNeedToMoveToSingleStep = new HashSet <string>();
needToMoveToSingleStepQueue.Enqueue(localNeedToMoveToSingleStep);
var localAlreadySingleStep = new HashSet <string>();
alreadySingleStepQueue.Enqueue(localAlreadySingleStep);
transactionalStorage.Batch(actions =>
{
var getItemsToReduceParams = new GetItemsToReduceParams(index: index.IndexId, reduceKeys: new HashSet <string>(localKeys), level: 0, loadData: false, itemsToDelete: itemsToDelete)
{
Take = int.MaxValue // just get all, we do the rate limit when we load the number of keys to reduce, anyway
};
var getItemsToReduceDuration = new Stopwatch();
int scheduledItemsSum = 0;
int scheduledItemsCount = 0;
List <int> scheduledItemsMappedBuckets = new List <int>();
using (StopwatchScope.For(getItemsToReduceDuration))
{
foreach (var item in actions.MapReduce.GetItemsToReduce(getItemsToReduceParams, token))
{
scheduledItemsMappedBuckets.Add(item.Bucket);
scheduledItemsSum += item.Size;
scheduledItemsCount++;
}
}
parallelStats.Operations.Add(PerformanceStats.From(IndexingOperation.Reduce_GetItemsToReduce, getItemsToReduceDuration.ElapsedMilliseconds));
autoTuner.CurrentlyUsedBatchSizesInBytes.GetOrAdd(reducingBatchThrottlerId, scheduledItemsSum);
if (scheduledItemsCount == 0)
{
// Here we have an interesting issue. We have scheduled reductions, because GetReduceTypesPerKeys() returned them
// and at the same time, we don't have any at level 0. That probably means that we have them at level 1 or 2.
// They shouldn't be here, and indeed, we remove them just a little down from here in this function.
// That said, they might have smuggled in between versions, or something happened to cause them to be here.
// In order to avoid that, we forcibly delete those extra items from the scheduled reductions, and move on
Log.Warn("Found single reduce items ({0}) that didn't have any items to reduce. Deleting level 1 & level 2 items for those keys. (If you can reproduce this, please contact [email protected])", string.Join(", ", keysToReduce));
var deletingScheduledReductionsDuration = Stopwatch.StartNew();
using (StopwatchScope.For(deletingScheduledReductionsDuration))
{
foreach (var reduceKey in keysToReduce)
{
token.ThrowIfCancellationRequested();
actions.MapReduce.DeleteScheduledReduction(index.IndexId, 1, reduceKey);
actions.MapReduce.DeleteScheduledReduction(index.IndexId, 2, reduceKey);
}
}
parallelStats.Operations.Add(PerformanceStats.From(IndexingOperation.Reduce_DeleteScheduledReductions, deletingScheduledReductionsDuration.ElapsedMilliseconds));
}
var removeReduceResultsDuration = new Stopwatch();
foreach (var reduceKey in localKeys)
{
token.ThrowIfCancellationRequested();
var lastPerformedReduceType = actions.MapReduce.GetLastPerformedReduceType(index.IndexId, reduceKey);
if (lastPerformedReduceType != ReduceType.SingleStep)
{
localNeedToMoveToSingleStep.Add(reduceKey);
}
if (lastPerformedReduceType == ReduceType.SingleStep)
{
localAlreadySingleStep.Add(reduceKey);
}
if (lastPerformedReduceType != ReduceType.MultiStep)
{
continue;
}
if (Log.IsDebugEnabled)
{
Log.Debug("Key {0} was moved from multi step to single step reduce, removing existing reduce results records", reduceKey);
}
using (StopwatchScope.For(removeReduceResultsDuration))
{
// now we are in single step but previously multi step reduce was performed for the given key
var mappedBuckets = actions.MapReduce.GetMappedBuckets(index.IndexId, reduceKey, token);
// add scheduled items too to be sure we will delete reduce results of already deleted documents
foreach (var mappedBucket in mappedBuckets.Union(scheduledItemsMappedBuckets))
{
actions.MapReduce.RemoveReduceResults(index.IndexId, 1, reduceKey, mappedBucket);
actions.MapReduce.RemoveReduceResults(index.IndexId, 2, reduceKey, mappedBucket / 1024);
}
}
}
parallelStats.Operations.Add(PerformanceStats.From(IndexingOperation.Reduce_RemoveReduceResults, removeReduceResultsDuration.ElapsedMilliseconds));
});
parallelOperations.Enqueue(parallelStats);
}
}, description: $"Performing single step reduction for index {index.Index.PublicName} from etag {index.Index.GetLastEtagFromStats()} for {keysToReduce.Count} keys", database: context.Database);
reduceLevelStats.Operations.Add(new ParallelPerformanceStats
{
NumberOfThreads = parallelOperations.Count,
DurationMs = (long)(SystemTime.UtcNow - parallelProcessingStart).TotalMilliseconds,
BatchedOperations = parallelOperations.ToList()
});
var getMappedResultsDuration = new Stopwatch();
var reductionPerformanceStats = new List <IndexingPerformanceStats>();
var keysLeftToReduce = new HashSet <string>(keysToReduce);
while (keysLeftToReduce.Count > 0)
{
var keysReturned = new HashSet <string>();
// Try to diminish the allocations happening because of .Resize()
var mappedResults = new List <MappedResultInfo>(keysLeftToReduce.Count);
context.TransactionalStorage.Batch(actions =>
{
var take = context.CurrentNumberOfItemsToReduceInSingleBatch;
using (StopwatchScope.For(getMappedResultsDuration))
{
mappedResults = actions.MapReduce.GetMappedResults(index.IndexId, keysLeftToReduce, true, take, keysReturned, token, mappedResults);
}
});
var count = mappedResults.Count;
int size = 0;
foreach (var item in mappedResults)
{
item.Bucket = 0;
size += item.Size;
}
var results = mappedResults.GroupBy(x => x.Bucket, x => JsonToExpando.Convert(x.Data)).ToArray();
context.MetricsCounters.ReducedPerSecond.Mark(results.Length);
token.ThrowIfCancellationRequested();
var performance = context.IndexStorage.Reduce(index.IndexId, viewGenerator, results, 2, context, null, keysReturned, count);
reductionPerformanceStats.Add(performance);
autoTuner.AutoThrottleBatchSize(count, size, batchTimeWatcher.Elapsed, skipIncreasingBatchSize);
}
// update new preformed single step
UpdatePerformedSingleStep(index.IndexId, needToMoveToSingleStepQueue);
// already single step,
// if the multi step keys were already removed,
// the reduce types for those keys needs to be removed also
UpdatePerformedSingleStep(index.IndexId, alreadySingleStepQueue, skipAdd: true);
reduceLevelStats.Completed = SystemTime.UtcNow;
reduceLevelStats.Duration = reduceLevelStats.Completed - reduceLevelStats.Started;
reduceLevelStats.Operations.Add(PerformanceStats.From(IndexingOperation.Reduce_GetMappedResults, getMappedResultsDuration.ElapsedMilliseconds));
reduceLevelStats.Operations.Add(PerformanceStats.From(IndexingOperation.StorageCommit, 0)); // in single step we write directly to Lucene index
foreach (var stats in reductionPerformanceStats)
{
reduceLevelStats.Add(stats);
}
reducePerformanceStats.LevelStats.Add(reduceLevelStats);
}
finally
{
long _;
autoTuner.CurrentlyUsedBatchSizesInBytes.TryRemove(reducingBatchThrottlerId, out _);
}
return(reducePerformanceStats);
}
public override void IndexDocuments(
AbstractViewGenerator viewGenerator,
IndexingBatch batch,
WorkContext context,
IStorageActionsAccessor actions,
DateTime minimumTimestamp)
{
var count = 0;
var sourceCount = 0;
var sw = Stopwatch.StartNew();
var start = SystemTime.UtcNow;
var changed = new HashSet <ReduceKeyAndBucket>();
var documentsWrapped = batch.Docs.Select(doc =>
{
sourceCount++;
var documentId = doc.__document_id;
actions.MapReduce.DeleteMappedResultsForDocumentId((string)documentId, name, changed);
return(doc);
})
.Where(x => x is FilteredDocument == false);
var stats = new IndexingWorkStats();
foreach (
var mappedResultFromDocument in
GroupByDocumentId(context,
RobustEnumerationIndex(documentsWrapped.GetEnumerator(), viewGenerator.MapDefinitions, actions, stats)))
{
var dynamicResults = mappedResultFromDocument.Select(x => (object)new DynamicJsonObject(RavenJObject.FromObject(x, jsonSerializer))).ToList();
foreach (
var doc in
RobustEnumerationReduceDuringMapPhase(dynamicResults.GetEnumerator(), viewGenerator.ReduceDefinition, actions, context))
{
count++;
var reduceValue = viewGenerator.GroupByExtraction(doc);
if (reduceValue == null)
{
logIndexing.Debug("Field {0} is used as the reduce key and cannot be null, skipping document {1}",
viewGenerator.GroupByExtraction, mappedResultFromDocument.Key);
continue;
}
var reduceKey = ReduceKeyToString(reduceValue);
var docId = mappedResultFromDocument.Key.ToString();
var data = GetMappedData(doc);
logIndexing.Debug("Mapped result for index '{0}' doc '{1}': '{2}'", name, docId, data);
actions.MapReduce.PutMappedResult(name, docId, reduceKey, data);
changed.Add(new ReduceKeyAndBucket(IndexingUtil.MapBucket(docId), reduceKey));
}
}
UpdateIndexingStats(context, stats);
actions.MapReduce.ScheduleReductions(name, 0, changed);
AddindexingPerformanceStat(new IndexingPerformanceStats
{
OutputCount = count,
InputCount = sourceCount,
Operation = "Map",
Duration = sw.Elapsed,
Started = start
});
logIndexing.Debug("Mapped {0} documents for {1}", count, name);
}
public static Sort GetSort(this IndexQuery self, IndexDefinition indexDefinition, AbstractViewGenerator viewGenerator)
{
var spatialQuery = self as SpatialIndexQuery;
var sortedFields = self.SortedFields;
if (sortedFields == null || sortedFields.Length <= 0)
{
if (spatialQuery == null || string.IsNullOrEmpty(self.Query) == false)
{
return(null);
}
sortedFields = new[] { new SortedField(Constants.DistanceFieldName), };
}
return(new Sort(sortedFields
.Select(sortedField =>
{
if (sortedField.Field == Constants.TemporaryScoreValue)
{
return SortField.FIELD_SCORE;
}
if (sortedField.Field.StartsWith(Constants.RandomFieldName))
{
var parts = sortedField.Field.Split(new[] { ';' }, StringSplitOptions.RemoveEmptyEntries);
if (parts.Length < 2) // truly random
{
return new RandomSortField(Guid.NewGuid().ToString());
}
return new RandomSortField(parts[1]);
}
if (spatialQuery != null && sortedField.Field == Constants.DistanceFieldName)
{
var spatialField = viewGenerator.GetSpatialField(spatialQuery.SpatialFieldName);
var shape = spatialField.ReadShape(spatialQuery.QueryShape);
var dsort = new SpatialDistanceFieldComparatorSource(spatialField, shape.GetCenter());
return new SortField(Constants.DistanceFieldName, dsort, sortedField.Descending);
}
var sortOptions = GetSortOption(indexDefinition, sortedField.Field, self);
if (sortOptions == null || sortOptions == SortOptions.None)
{
return new SortField(sortedField.Field, CultureInfo.InvariantCulture, sortedField.Descending);
}
if (sortOptions.Value == SortOptions.Short)
{
sortOptions = SortOptions.Int;
}
return new SortField(sortedField.Field, (int)sortOptions.Value, sortedField.Descending);
})
.ToArray()));
}
public override IndexingPerformanceStats IndexDocuments(AbstractViewGenerator viewGenerator, IndexingBatch batch, IStorageActionsAccessor actions, DateTime minimumTimestamp, CancellationToken token)
{
token.ThrowIfCancellationRequested();
var count = 0;
var sourceCount = 0;
var deleted = new Dictionary <ReduceKeyAndBucket, int>();
var performance = RecordCurrentBatch("Current Map", "Map", batch.Docs.Count);
var performanceStats = new List <BasePerformanceStats>();
var usedStorageAccessors = new ConcurrentSet <IStorageActionsAccessor>();
if (usedStorageAccessors.TryAdd(actions))
{
var storageCommitDuration = new Stopwatch();
actions.BeforeStorageCommit += storageCommitDuration.Start;
actions.AfterStorageCommit += () =>
{
storageCommitDuration.Stop();
performanceStats.Add(PerformanceStats.From(IndexingOperation.StorageCommit, storageCommitDuration.ElapsedMilliseconds));
};
}
var deleteMappedResultsDuration = new Stopwatch();
var documentsWrapped = batch.Docs.Select(doc =>
{
token.ThrowIfCancellationRequested();
sourceCount++;
var documentId = doc.__document_id;
using (StopwatchScope.For(deleteMappedResultsDuration))
{
actions.MapReduce.DeleteMappedResultsForDocumentId((string)documentId, indexId, deleted);
}
return(doc);
})
.Where(x => x is FilteredDocument == false)
.ToList();
performanceStats.Add(new PerformanceStats
{
Name = IndexingOperation.Map_DeleteMappedResults,
DurationMs = deleteMappedResultsDuration.ElapsedMilliseconds,
});
var allReferencedDocs = new ConcurrentQueue <IDictionary <string, HashSet <string> > >();
var allReferenceEtags = new ConcurrentQueue <IDictionary <string, Etag> >();
var allState = new ConcurrentQueue <Tuple <HashSet <ReduceKeyAndBucket>, IndexingWorkStats, Dictionary <string, int> > >();
var parallelOperations = new ConcurrentQueue <ParallelBatchStats>();
var parallelProcessingStart = SystemTime.UtcNow;
BackgroundTaskExecuter.Instance.ExecuteAllBuffered(context, documentsWrapped, partition =>
{
token.ThrowIfCancellationRequested();
var parallelStats = new ParallelBatchStats
{
StartDelay = (long)(SystemTime.UtcNow - parallelProcessingStart).TotalMilliseconds
};
var localStats = new IndexingWorkStats();
var localChanges = new HashSet <ReduceKeyAndBucket>();
var statsPerKey = new Dictionary <string, int>();
var linqExecutionDuration = new Stopwatch();
var reduceInMapLinqExecutionDuration = new Stopwatch();
var putMappedResultsDuration = new Stopwatch();
var convertToRavenJObjectDuration = new Stopwatch();
allState.Enqueue(Tuple.Create(localChanges, localStats, statsPerKey));
using (CurrentIndexingScope.Current = new CurrentIndexingScope(context.Database, PublicName))
{
// we are writing to the transactional store from multiple threads here, and in a streaming fashion
// should result in less memory and better perf
context.TransactionalStorage.Batch(accessor =>
{
if (usedStorageAccessors.TryAdd(accessor))
{
var storageCommitDuration = new Stopwatch();
accessor.BeforeStorageCommit += storageCommitDuration.Start;
accessor.AfterStorageCommit += () =>
{
storageCommitDuration.Stop();
parallelStats.Operations.Add(PerformanceStats.From(IndexingOperation.StorageCommit, storageCommitDuration.ElapsedMilliseconds));
};
}
var mapResults = RobustEnumerationIndex(partition, viewGenerator.MapDefinitions, localStats, linqExecutionDuration);
var currentDocumentResults = new List <object>();
string currentKey = null;
bool skipDocument = false;
foreach (var currentDoc in mapResults)
{
token.ThrowIfCancellationRequested();
var documentId = GetDocumentId(currentDoc);
if (documentId != currentKey)
{
count += ProcessBatch(viewGenerator, currentDocumentResults, currentKey, localChanges, accessor, statsPerKey, reduceInMapLinqExecutionDuration, putMappedResultsDuration, convertToRavenJObjectDuration);
currentDocumentResults.Clear();
currentKey = documentId;
}
else if (skipDocument)
{
continue;
}
RavenJObject currentDocJObject;
using (StopwatchScope.For(convertToRavenJObjectDuration))
{
currentDocJObject = RavenJObject.FromObject(currentDoc, jsonSerializer);
}
currentDocumentResults.Add(new DynamicJsonObject(currentDocJObject));
if (EnsureValidNumberOfOutputsForDocument(documentId, currentDocumentResults.Count) == false)
{
skipDocument = true;
currentDocumentResults.Clear();
continue;
}
Interlocked.Increment(ref localStats.IndexingSuccesses);
}
count += ProcessBatch(viewGenerator, currentDocumentResults, currentKey, localChanges, accessor, statsPerKey, reduceInMapLinqExecutionDuration, putMappedResultsDuration, convertToRavenJObjectDuration);
parallelStats.Operations.Add(PerformanceStats.From(IndexingOperation.LoadDocument, CurrentIndexingScope.Current.LoadDocumentDuration.ElapsedMilliseconds));
parallelStats.Operations.Add(PerformanceStats.From(IndexingOperation.Linq_MapExecution, linqExecutionDuration.ElapsedMilliseconds));
parallelStats.Operations.Add(PerformanceStats.From(IndexingOperation.Linq_ReduceLinqExecution, reduceInMapLinqExecutionDuration.ElapsedMilliseconds));
parallelStats.Operations.Add(PerformanceStats.From(IndexingOperation.Map_PutMappedResults, putMappedResultsDuration.ElapsedMilliseconds));
parallelStats.Operations.Add(PerformanceStats.From(IndexingOperation.Map_ConvertToRavenJObject, convertToRavenJObjectDuration.ElapsedMilliseconds));
parallelOperations.Enqueue(parallelStats);
});
allReferenceEtags.Enqueue(CurrentIndexingScope.Current.ReferencesEtags);
allReferencedDocs.Enqueue(CurrentIndexingScope.Current.ReferencedDocuments);
}
});
performanceStats.Add(new ParallelPerformanceStats
{
NumberOfThreads = parallelOperations.Count,
DurationMs = (long)(SystemTime.UtcNow - parallelProcessingStart).TotalMilliseconds,
BatchedOperations = parallelOperations.ToList()
});
var updateDocumentReferencesDuration = new Stopwatch();
using (StopwatchScope.For(updateDocumentReferencesDuration))
{
UpdateDocumentReferences(actions, allReferencedDocs, allReferenceEtags);
}
performanceStats.Add(PerformanceStats.From(IndexingOperation.UpdateDocumentReferences, updateDocumentReferencesDuration.ElapsedMilliseconds));
var changed = allState.SelectMany(x => x.Item1).Concat(deleted.Keys)
.Distinct()
.ToList();
var stats = new IndexingWorkStats(allState.Select(x => x.Item2));
var reduceKeyStats = allState.SelectMany(x => x.Item3)
.GroupBy(x => x.Key)
.Select(g => new { g.Key, Count = g.Sum(x => x.Value) })
.ToList();
BackgroundTaskExecuter.Instance.ExecuteAllBuffered(context, reduceKeyStats, enumerator => context.TransactionalStorage.Batch(accessor =>
{
while (enumerator.MoveNext())
{
var reduceKeyStat = enumerator.Current;
accessor.MapReduce.IncrementReduceKeyCounter(indexId, reduceKeyStat.Key, reduceKeyStat.Count);
}
}));
actions.General.MaybePulseTransaction();
var parallelReductionOperations = new ConcurrentQueue <ParallelBatchStats>();
var parallelReductionStart = SystemTime.UtcNow;
BackgroundTaskExecuter.Instance.ExecuteAllBuffered(context, changed, enumerator => context.TransactionalStorage.Batch(accessor =>
{
var parallelStats = new ParallelBatchStats
{
StartDelay = (long)(SystemTime.UtcNow - parallelReductionStart).TotalMilliseconds
};
var scheduleReductionsDuration = new Stopwatch();
using (StopwatchScope.For(scheduleReductionsDuration))
{
while (enumerator.MoveNext())
{
accessor.MapReduce.ScheduleReductions(indexId, 0, enumerator.Current);
accessor.General.MaybePulseTransaction();
}
}
parallelStats.Operations.Add(PerformanceStats.From(IndexingOperation.Map_ScheduleReductions, scheduleReductionsDuration.ElapsedMilliseconds));
parallelReductionOperations.Enqueue(parallelStats);
}));
performanceStats.Add(new ParallelPerformanceStats
{
NumberOfThreads = parallelReductionOperations.Count,
DurationMs = (long)(SystemTime.UtcNow - parallelReductionStart).TotalMilliseconds,
BatchedOperations = parallelReductionOperations.ToList()
});
UpdateIndexingStats(context, stats);
performance.OnCompleted = () => BatchCompleted("Current Map", "Map", sourceCount, count, performanceStats);
logIndexing.Debug("Mapped {0} documents for {1}", count, PublicName);
return(performance);
}
private int ProcessBatch(AbstractViewGenerator viewGenerator, List <object> currentDocumentResults, string currentKey, HashSet <ReduceKeyAndBucket> changes,
IStorageActionsAccessor actions,
IDictionary <string, int> statsPerKey, Stopwatch reduceDuringMapLinqExecution, Stopwatch putMappedResultsDuration, Stopwatch convertToRavenJObjectDuration)
{
if (currentKey == null || currentDocumentResults.Count == 0)
{
return(0);
}
var old = CurrentIndexingScope.Current;
try
{
CurrentIndexingScope.Current = null;
if (logIndexing.IsDebugEnabled)
{
var sb = new StringBuilder()
.AppendFormat("Index {0} for document {1} resulted in:", PublicName, currentKey)
.AppendLine();
foreach (var currentDocumentResult in currentDocumentResults)
{
sb.AppendLine(JsonConvert.SerializeObject(currentDocumentResult));
}
logIndexing.Debug(sb.ToString());
}
int count = 0;
var results = RobustEnumerationReduceDuringMapPhase(currentDocumentResults.GetEnumerator(), viewGenerator.ReduceDefinition, reduceDuringMapLinqExecution);
foreach (var doc in results)
{
count++;
var reduceValue = viewGenerator.GroupByExtraction(doc);
if (reduceValue == null)
{
logIndexing.Debug("Field {0} is used as the reduce key and cannot be null, skipping document {1}",
viewGenerator.GroupByExtraction, currentKey);
continue;
}
string reduceKey = ReduceKeyToString(reduceValue);
RavenJObject data;
using (StopwatchScope.For(convertToRavenJObjectDuration))
{
data = GetMappedData(doc);
}
if (logIndexing.IsDebugEnabled)
{
logIndexing.Debug("Index {0} for document {1} resulted in ({2}): {3}", PublicName, currentKey, reduceKey, data);
}
using (StopwatchScope.For(putMappedResultsDuration))
{
actions.MapReduce.PutMappedResult(indexId, currentKey, reduceKey, data);
}
statsPerKey[reduceKey] = statsPerKey.GetOrDefault(reduceKey) + 1;
actions.General.MaybePulseTransaction();
changes.Add(new ReduceKeyAndBucket(IndexingUtil.MapBucket(currentKey), reduceKey));
}
return(count);
}
finally
{
CurrentIndexingScope.Current = old;
}
}
private void ApplyPrecomputedBatchForNewIndex(Index index, AbstractViewGenerator generator)
{
PrecomputedIndexingBatch result = null;
var docsToIndex = new List <JsonDocument>();
TransactionalStorage.Batch(actions =>
{
var query = GetQueryForAllMatchingDocumentsForIndex(generator);
JsonDocument highestByEtag = null;
var cts = new CancellationTokenSource();
using (var linked = CancellationTokenSource.CreateLinkedTokenSource(cts.Token, WorkContext.CancellationToken))
using (var op = new QueryActions.DatabaseQueryOperation(Database, Constants.DocumentsByEntityNameIndex, new IndexQuery
{
Query = query,
PageSize = Database.Configuration.MaxNumberOfItemsToProcessInSingleBatch
}, actions, linked)
{
ShouldSkipDuplicateChecking = true
})
{
op.Init();
if (op.Header.TotalResults == 0 ||
(op.Header.TotalResults > Database.Configuration.MaxNumberOfItemsToProcessInSingleBatch))
{
// we don't apply this optimization if the total number of results
// to index is more than the max numbers to index in a single batch.
// The idea here is that we need to keep the amount
// of memory we use to a manageable level even when introducing a new index to a BIG
// database
try
{
cts.Cancel();
// we have to run just a little bit of the query to properly setup the disposal
op.Execute(o => { });
}
catch (OperationCanceledException)
{
}
return;
}
Log.Debug("For new index {0}, using precomputed indexing batch optimization for {1} docs", index,
op.Header.TotalResults);
op.Execute(document =>
{
var metadata = document.Value <RavenJObject>(Constants.Metadata);
var key = metadata.Value <string>("@id");
var etag = Etag.Parse(metadata.Value <string>("@etag"));
var lastModified = DateTime.Parse(metadata.Value <string>(Constants.LastModified));
document.Remove(Constants.Metadata);
var doc = new JsonDocument
{
DataAsJson = document,
Etag = etag,
Key = key,
LastModified = lastModified,
SkipDeleteFromIndex = true,
Metadata = metadata
};
docsToIndex.Add(doc);
if (highestByEtag == null || doc.Etag.CompareTo(highestByEtag.Etag) > 0)
{
highestByEtag = doc;
}
});
}
result = new PrecomputedIndexingBatch
{
LastIndexed = highestByEtag.Etag,
LastModified = highestByEtag.LastModified.Value,
Documents = docsToIndex,
Index = index
};
});
if (result != null && result.Documents != null && result.Documents.Count > 0)
{
Database.IndexingExecuter.IndexPrecomputedBatch(result);
}
}
