// Delete by query
private static long ApplyQueryDeletes(IEnumerable <QueryAndLimit> queriesIter, ReadersAndUpdates rld, SegmentReader reader)
{
long delCount = 0;
AtomicReaderContext readerContext = reader.AtomicContext;
bool any = false;
foreach (QueryAndLimit ent in queriesIter)
{
Query query = ent.Query;
int? limit = ent.Limit;
DocIdSet docs = (new QueryWrapperFilter(query)).GetDocIdSet(readerContext, reader.LiveDocs);
if (docs != null)
{
DocIdSetIterator it = docs.GetIterator();
if (it != null)
{
while (true)
{
int doc = it.NextDoc();
if (doc >= limit)
{
break;
}
if (!any)
{
rld.InitWritableLiveDocs();
any = true;
}
if (rld.Delete(doc))
{
delCount++;
}
}
}
}
}
return(delCount);
}
/// <summary>
/// Used by near real-time search </summary>
internal static DirectoryReader Open(IndexWriter writer, SegmentInfos infos, bool applyAllDeletes)
{
// IndexWriter synchronizes externally before calling
// us, which ensures infos will not change; so there's
// no need to process segments in reverse order
int numSegments = infos.Count;
IList <SegmentReader> readers = new List <SegmentReader>();
Directory dir = writer.Directory;
SegmentInfos segmentInfos = (SegmentInfos)infos.Clone();
int infosUpto = 0;
bool success = false;
try
{
for (int i = 0; i < numSegments; i++)
{
// NOTE: important that we use infos not
// segmentInfos here, so that we are passing the
// actual instance of SegmentInfoPerCommit in
// IndexWriter's segmentInfos:
SegmentCommitInfo info = infos.Info(i);
Debug.Assert(info.Info.Dir == dir);
ReadersAndUpdates rld = writer.readerPool.Get(info, true);
try
{
SegmentReader reader = rld.GetReadOnlyClone(IOContext.READ);
if (reader.NumDocs > 0 || writer.KeepFullyDeletedSegments)
{
// Steal the ref:
readers.Add(reader);
infosUpto++;
}
else
{
reader.DecRef();
segmentInfos.Remove(infosUpto);
}
}
finally
{
writer.readerPool.Release(rld);
}
}
writer.IncRefDeleter(segmentInfos);
StandardDirectoryReader result = new StandardDirectoryReader(dir, readers.ToArray(), writer, segmentInfos, writer.Config.ReaderTermsIndexDivisor, applyAllDeletes);
success = true;
return(result);
}
finally
{
if (!success)
{
foreach (SegmentReader r in readers)
{
try
{
r.DecRef();
}
#pragma warning disable 168
catch (Exception th)
#pragma warning restore 168
{
// ignore any exception that is thrown here to not mask any original
// exception.
}
}
}
}
}
// Delete by Term
private long ApplyTermDeletes(IEnumerable <Term> termsIter, ReadersAndUpdates rld, SegmentReader reader)
{
lock (this)
{
long delCount = 0;
Fields fields = reader.Fields;
if (fields == null)
{
// this reader has no postings
return(0);
}
TermsEnum termsEnum = null;
string currentField = null;
DocsEnum docs = null;
Debug.Assert(CheckDeleteTerm(null));
bool any = false;
//System.out.println(Thread.currentThread().getName() + " del terms reader=" + reader);
foreach (Term term in termsIter)
{
// Since we visit terms sorted, we gain performance
// by re-using the same TermsEnum and seeking only
// forwards
if (!term.Field().Equals(currentField))
{
Debug.Assert(currentField == null || currentField.CompareTo(term.Field()) < 0);
currentField = term.Field();
Terms terms = fields.Terms(currentField);
if (terms != null)
{
termsEnum = terms.Iterator(termsEnum);
}
else
{
termsEnum = null;
}
}
if (termsEnum == null)
{
continue;
}
Debug.Assert(CheckDeleteTerm(term));
// System.out.println(" term=" + term);
if (termsEnum.SeekExact(term.Bytes()))
{
// we don't need term frequencies for this
DocsEnum docsEnum = termsEnum.Docs(rld.LiveDocs, docs, DocsEnum.FLAG_NONE);
//System.out.println("BDS: got docsEnum=" + docsEnum);
if (docsEnum != null)
{
while (true)
{
int docID = docsEnum.NextDoc();
//System.out.println(Thread.currentThread().getName() + " del term=" + term + " doc=" + docID);
if (docID == DocIdSetIterator.NO_MORE_DOCS)
{
break;
}
if (!any)
{
rld.InitWritableLiveDocs();
any = true;
}
// NOTE: there is no limit check on the docID
// when deleting by Term (unlike by Query)
// because on flush we apply all Term deletes to
// each segment. So all Term deleting here is
// against prior segments:
if (rld.Delete(docID))
{
delCount++;
}
}
}
}
}
return(delCount);
}
}
// DocValues updates
private void ApplyDocValuesUpdates <T1>(IEnumerable <T1> updates, ReadersAndUpdates rld, SegmentReader reader, DocValuesFieldUpdates.Container dvUpdatesContainer) where T1 : DocValuesUpdate
{
lock (this)
{
Fields fields = reader.Fields;
if (fields == null)
{
// this reader has no postings
return;
}
// TODO: we can process the updates per DV field, from last to first so that
// if multiple terms affect same document for the same field, we add an update
// only once (that of the last term). To do that, we can keep a bitset which
// marks which documents have already been updated. So e.g. if term T1
// updates doc 7, and then we process term T2 and it updates doc 7 as well,
// we don't apply the update since we know T1 came last and therefore wins
// the update.
// We can also use that bitset as 'liveDocs' to pass to TermEnum.docs(), so
// that these documents aren't even returned.
string currentField = null;
TermsEnum termsEnum = null;
DocsEnum docs = null;
//System.out.println(Thread.currentThread().getName() + " numericDVUpdate reader=" + reader);
foreach (DocValuesUpdate update in updates)
{
Term term = update.Term;
int limit = update.DocIDUpto;
// TODO: we traverse the terms in update order (not term order) so that we
// apply the updates in the correct order, i.e. if two terms udpate the
// same document, the last one that came in wins, irrespective of the
// terms lexical order.
// we can apply the updates in terms order if we keep an updatesGen (and
// increment it with every update) and attach it to each NumericUpdate. Note
// that we cannot rely only on docIDUpto because an app may send two updates
// which will get same docIDUpto, yet will still need to respect the order
// those updates arrived.
if (!term.Field().Equals(currentField))
{
// if we change the code to process updates in terms order, enable this assert
// assert currentField == null || currentField.compareTo(term.field()) < 0;
currentField = term.Field();
Terms terms = fields.Terms(currentField);
if (terms != null)
{
termsEnum = terms.Iterator(termsEnum);
}
else
{
termsEnum = null;
continue; // no terms in that field
}
}
if (termsEnum == null)
{
continue;
}
// System.out.println(" term=" + term);
if (termsEnum.SeekExact(term.Bytes()))
{
// we don't need term frequencies for this
DocsEnum docsEnum = termsEnum.Docs(rld.LiveDocs, docs, DocsEnum.FLAG_NONE);
//System.out.println("BDS: got docsEnum=" + docsEnum);
DocValuesFieldUpdates dvUpdates = dvUpdatesContainer.GetUpdates(update.Field, update.Type);
if (dvUpdates == null)
{
dvUpdates = dvUpdatesContainer.NewUpdates(update.Field, update.Type, reader.MaxDoc);
}
int doc;
while ((doc = docsEnum.NextDoc()) != DocIdSetIterator.NO_MORE_DOCS)
{
//System.out.println(Thread.currentThread().getName() + " numericDVUpdate term=" + term + " doc=" + docID);
if (doc >= limit)
{
break; // no more docs that can be updated for this term
}
dvUpdates.Add(doc, update.Value);
}
}
}
}
}
/// <summary>
/// Resolves the buffered deleted Term/Query/docIDs, into
/// actual deleted docIDs in the liveDocs MutableBits for
/// each SegmentReader.
/// </summary>
public virtual ApplyDeletesResult ApplyDeletesAndUpdates(IndexWriter.ReaderPool readerPool, IList <SegmentCommitInfo> infos)
{
lock (this)
{
long t0 = DateTime.Now.Millisecond;
if (infos.Count == 0)
{
return(new ApplyDeletesResult(false, NextGen_Renamed++, null));
}
Debug.Assert(CheckDeleteStats());
if (!Any())
{
if (InfoStream.IsEnabled("BD"))
{
InfoStream.Message("BD", "applyDeletes: no deletes; skipping");
}
return(new ApplyDeletesResult(false, NextGen_Renamed++, null));
}
if (InfoStream.IsEnabled("BD"))
{
InfoStream.Message("BD", "applyDeletes: infos=" + infos + " packetCount=" + Updates.Count);
}
long gen = NextGen_Renamed++;
List <SegmentCommitInfo> infos2 = new List <SegmentCommitInfo>();
infos2.AddRange(infos);
infos2.Sort(sortSegInfoByDelGen);
CoalescedUpdates coalescedUpdates = null;
bool anyNewDeletes = false;
int infosIDX = infos2.Count - 1;
int delIDX = Updates.Count - 1;
IList <SegmentCommitInfo> allDeleted = null;
while (infosIDX >= 0)
{
//System.out.println("BD: cycle delIDX=" + delIDX + " infoIDX=" + infosIDX);
FrozenBufferedUpdates packet = delIDX >= 0 ? Updates[delIDX] : null;
SegmentCommitInfo info = infos2[infosIDX];
long segGen = info.BufferedDeletesGen;
if (packet != null && segGen < packet.DelGen)
{
// System.out.println(" coalesce");
if (coalescedUpdates == null)
{
coalescedUpdates = new CoalescedUpdates();
}
if (!packet.IsSegmentPrivate)
{
/*
* Only coalesce if we are NOT on a segment private del packet: the segment private del packet
* must only applied to segments with the same delGen. Yet, if a segment is already deleted
* from the SI since it had no more documents remaining after some del packets younger than
* its segPrivate packet (higher delGen) have been applied, the segPrivate packet has not been
* removed.
*/
coalescedUpdates.Update(packet);
}
delIDX--;
}
else if (packet != null && segGen == packet.DelGen)
{
Debug.Assert(packet.IsSegmentPrivate, "Packet and Segments deletegen can only match on a segment private del packet gen=" + segGen);
//System.out.println(" eq");
// Lock order: IW -> BD -> RP
Debug.Assert(readerPool.InfoIsLive(info));
ReadersAndUpdates rld = readerPool.Get(info, true);
SegmentReader reader = rld.GetReader(IOContext.READ);
int delCount = 0;
bool segAllDeletes;
try
{
DocValuesFieldUpdates.Container dvUpdates = new DocValuesFieldUpdates.Container();
if (coalescedUpdates != null)
{
//System.out.println(" del coalesced");
delCount += (int)ApplyTermDeletes(coalescedUpdates.TermsIterable(), rld, reader);
delCount += (int)ApplyQueryDeletes(coalescedUpdates.QueriesIterable(), rld, reader);
ApplyDocValuesUpdates(coalescedUpdates.NumericDVUpdates, rld, reader, dvUpdates);
ApplyDocValuesUpdates(coalescedUpdates.BinaryDVUpdates, rld, reader, dvUpdates);
}
//System.out.println(" del exact");
// Don't delete by Term here; DocumentsWriterPerThread
// already did that on flush:
delCount += (int)ApplyQueryDeletes(packet.QueriesIterable(), rld, reader);
ApplyDocValuesUpdates(Arrays.AsList(packet.NumericDVUpdates), rld, reader, dvUpdates);
ApplyDocValuesUpdates(Arrays.AsList(packet.BinaryDVUpdates), rld, reader, dvUpdates);
if (dvUpdates.Any())
{
rld.WriteFieldUpdates(info.Info.Dir, dvUpdates);
}
int fullDelCount = rld.Info.DelCount + rld.PendingDeleteCount;
Debug.Assert(fullDelCount <= rld.Info.Info.DocCount);
segAllDeletes = fullDelCount == rld.Info.Info.DocCount;
}
finally
{
rld.Release(reader);
readerPool.Release(rld);
}
anyNewDeletes |= delCount > 0;
if (segAllDeletes)
{
if (allDeleted == null)
{
allDeleted = new List <SegmentCommitInfo>();
}
allDeleted.Add(info);
}
if (InfoStream.IsEnabled("BD"))
{
InfoStream.Message("BD", "seg=" + info + " segGen=" + segGen + " segDeletes=[" + packet + "]; coalesced deletes=[" + (coalescedUpdates == null ? "null" : coalescedUpdates.ToString()) + "] newDelCount=" + delCount + (segAllDeletes ? " 100% deleted" : ""));
}
if (coalescedUpdates == null)
{
coalescedUpdates = new CoalescedUpdates();
}
/*
* Since we are on a segment private del packet we must not
* update the coalescedDeletes here! We can simply advance to the
* next packet and seginfo.
*/
delIDX--;
infosIDX--;
info.BufferedDeletesGen = gen;
}
else
{
//System.out.println(" gt");
if (coalescedUpdates != null)
{
// Lock order: IW -> BD -> RP
Debug.Assert(readerPool.InfoIsLive(info));
ReadersAndUpdates rld = readerPool.Get(info, true);
SegmentReader reader = rld.GetReader(IOContext.READ);
int delCount = 0;
bool segAllDeletes;
try
{
delCount += (int)ApplyTermDeletes(coalescedUpdates.TermsIterable(), rld, reader);
delCount += (int)ApplyQueryDeletes(coalescedUpdates.QueriesIterable(), rld, reader);
DocValuesFieldUpdates.Container dvUpdates = new DocValuesFieldUpdates.Container();
ApplyDocValuesUpdates(coalescedUpdates.NumericDVUpdates, rld, reader, dvUpdates);
ApplyDocValuesUpdates(coalescedUpdates.BinaryDVUpdates, rld, reader, dvUpdates);
if (dvUpdates.Any())
{
rld.WriteFieldUpdates(info.Info.Dir, dvUpdates);
}
int fullDelCount = rld.Info.DelCount + rld.PendingDeleteCount;
Debug.Assert(fullDelCount <= rld.Info.Info.DocCount);
segAllDeletes = fullDelCount == rld.Info.Info.DocCount;
}
finally
{
rld.Release(reader);
readerPool.Release(rld);
}
anyNewDeletes |= delCount > 0;
if (segAllDeletes)
{
if (allDeleted == null)
{
allDeleted = new List <SegmentCommitInfo>();
}
allDeleted.Add(info);
}
if (InfoStream.IsEnabled("BD"))
{
InfoStream.Message("BD", "seg=" + info + " segGen=" + segGen + " coalesced deletes=[" + coalescedUpdates + "] newDelCount=" + delCount + (segAllDeletes ? " 100% deleted" : ""));
}
}
info.BufferedDeletesGen = gen;
infosIDX--;
}
}
Debug.Assert(CheckDeleteStats());
if (InfoStream.IsEnabled("BD"))
{
InfoStream.Message("BD", "applyDeletes took " + (DateTime.Now.Millisecond - t0) + " msec");
}
// assert infos != segmentInfos || !any() : "infos=" + infos + " segmentInfos=" + segmentInfos + " any=" + any;
return(new ApplyDeletesResult(anyNewDeletes, gen, allDeleted));
}
}
/// <summary>
/// Obtain a ReadersAndLiveDocs instance from the
/// readerPool. If create is true, you must later call
/// <seealso cref="#release(ReadersAndUpdates)"/>.
/// </summary>
public virtual ReadersAndUpdates Get(SegmentCommitInfo info, bool create)
{
lock (this)
{
Debug.Assert(info.Info.Dir == OuterInstance.directory, "info.dir=" + info.Info.Dir + " vs " + OuterInstance.directory);
ReadersAndUpdates rld;
ReaderMap.TryGetValue(info, out rld);
if (rld == null)
{
if (!create)
{
return null;
}
rld = new ReadersAndUpdates(OuterInstance, info);
// Steal initial reference:
ReaderMap[info] = rld;
}
else
{
Debug.Assert(rld.Info == info, "Infos are not equal");//, "rld.info=" + rld.Info + " info=" + info + " isLive?=" + InfoIsLive(rld.Info) + " vs " + InfoIsLive(info));
}
if (create)
{
// Return ref to caller:
rld.IncRef();
}
Debug.Assert(NoDups());
return rld;
}
}
public virtual void Release(ReadersAndUpdates rld, bool assertInfoLive)
{
lock (this)
{
// Matches incRef in get:
rld.DecRef();
// Pool still holds a ref:
Debug.Assert(rld.RefCount() >= 1);
if (!OuterInstance.PoolReaders && rld.RefCount() == 1)
{
// this is the last ref to this RLD, and we're not
// pooling, so remove it:
// System.out.println("[" + Thread.currentThread().getName() + "] ReaderPool.release: " + rld.info);
if (rld.WriteLiveDocs(OuterInstance.directory))
{
// Make sure we only write del docs for a live segment:
Debug.Assert(assertInfoLive == false || InfoIsLive(rld.Info));
// Must checkpoint because we just
// created new _X_N.del and field updates files;
// don't call IW.checkpoint because that also
// increments SIS.version, which we do not want to
// do here: it was done previously (after we
// invoked BDS.applyDeletes), whereas here all we
// did was move the state to disk:
OuterInstance.CheckpointNoSIS();
}
//System.out.println("IW: done writeLiveDocs for info=" + rld.info);
// System.out.println("[" + Thread.currentThread().getName() + "] ReaderPool.release: drop readers " + rld.info);
rld.DropReaders();
ReaderMap.Remove(rld.Info);
}
}
}
public virtual void Release(ReadersAndUpdates rld)
{
lock (this)
{
Release(rld, true);
}
}
internal void Init(ReaderPool readerPool, MergePolicy.OneMerge merge, MergeState mergeState, bool initWritableLiveDocs)
{
if (MergedDeletesAndUpdates_Renamed == null)
{
MergedDeletesAndUpdates_Renamed = readerPool.Get(merge.Info_Renamed, true);
DocMap = merge.GetDocMap(mergeState);
Debug.Assert(DocMap.IsConsistent(merge.Info_Renamed.Info.DocCount));
}
if (initWritableLiveDocs && !InitializedWritableLiveDocs)
{
MergedDeletesAndUpdates_Renamed.InitWritableLiveDocs();
this.InitializedWritableLiveDocs = true;
}
}
