From 2418da510eaf2fc5d1661ba816942086dfc9090e Mon Sep 17 00:00:00 2001
From: Eliot Horowitz <eliot@10gen.com>
Date: Sun, 12 Sep 2010 18:23:27 -0400
Subject: [PATCH] use splitVector instead of datasize and mediankey.  1 command
 vs. 2 and should be faster and more extensible. still some code cleanup to do
 probably SERVER-1550 SERVER-1779

---
 s/chunk.cpp   | 37 +++++++++++++++++++++++++++----------
 s/chunk.h     | 10 +++++++++-
 s/d_split.cpp | 25 ++++++++++++++++++-------
 3 files changed, 54 insertions(+), 18 deletions(-)

diff --git a/s/chunk.cpp b/s/chunk.cpp
index dafa42509fb..3f980ba6575 100644
--- a/s/chunk.cpp
+++ b/s/chunk.cpp
@@ -83,7 +83,7 @@ namespace mongo {
         return _manager->getShardKey().globalMax().woCompare( getMax() ) == 0;
     }
     
-    BSONObj Chunk::pickSplitPoint() const{
+    BSONObj Chunk::pickSplitPoint( const vector<BSONObj> * possibleSplitPoints ) const{
         int sort = 0;
         
         if ( minIsInf() ){
@@ -117,6 +117,9 @@ namespace mongo {
             if ( ! end.isEmpty() )
                 return _manager->getShardKey().extractKey( end );
         }
+
+        if ( possibleSplitPoints && possibleSplitPoints->size() )
+            return possibleSplitPoints->at(0);
         
         BSONObj cmd = BSON( "medianKey" << _manager->getns()
                             << "keyPattern" << _manager->getShardKey().key()
@@ -164,7 +167,7 @@ namespace mongo {
         cmd.append( "keyPattern" , _manager->getShardKey().key() );
         cmd.append( "min" , getMin() );
         cmd.append( "max" , getMax() );
-        cmd.append( "maxChunkSize" , chunkSize / (1<<20) /* in MBs */ );
+        cmd.append( "maxChunkSizeBytes" , chunkSize );
         BSONObj cmdObj = cmd.obj();
 
         if ( ! conn->runCommand( "admin" , cmdObj , result )){
@@ -377,21 +380,35 @@ namespace mongo {
             
             log(3) << "\t splitIfShould entering decision area : " << *this << endl;
             
-            _dataWritten = 0;
+            _dataWritten = 0; // reset so we check often enough
             
-            BSONObj splitPoint = pickSplitPoint();
+            // TODO: add a max number of split points to find
+            //       that way if we get a mega chunk for some reason - 
+            //       this won't take an inordinant amount of time
+            vector<BSONObj> possibleSplitPoints;
+            pickSplitVector( possibleSplitPoints , splitThreshold );
+            
+            if ( possibleSplitPoints.size() <= 1 ) {
+                // no split points means there isn't enough data to split on
+                // 1 split point means we have between half the chunk size to full chunk size
+                // so we shouldn't split
+                return false;
+            }
+
+            BSONObj splitPoint = pickSplitPoint( &possibleSplitPoints );
             if ( splitPoint.isEmpty() || _min == splitPoint || _max == splitPoint) {
+                // TODO: this check might be redundany, but probably not that bad
                 error() << "want to split chunk, but can't find split point " 
                         << " chunk: " << toString() << " got: " << splitPoint << endl;
                 return false;
             }
             
-            long size = getPhysicalSize();
-            if ( size < splitThreshold )
-                return false;
-            
-            log() << "autosplitting " << _manager->getns() << " size: " << size << " shard: " << toString() 
-                  << " on: " << splitPoint << "(splitThreshold " << splitThreshold << ")" << endl;
+            log() << "autosplitting " << _manager->getns() << " shard: " << toString() 
+                  << " on: " << splitPoint << "(splitThreshold " << splitThreshold << ")" 
+#ifdef _DEBUG
+                  << " size: " << getPhysicalSize() // slow - but can be usefule when debugging
+#endif
+                  << endl;
             
             vector<BSONObj> splitPoints;
             splitPoints.push_back( splitPoint );
diff --git a/s/chunk.h b/s/chunk.h
index 0f23a8ed23e..6e5ef591d12 100644
--- a/s/chunk.h
+++ b/s/chunk.h
@@ -88,9 +88,17 @@ namespace mongo {
         bool minIsInf() const;
         bool maxIsInf() const;
 
-        BSONObj pickSplitPoint() const;
+        /**
+         * @param a vector of possible split points
+         *        used as a hint only
+         */
+        BSONObj pickSplitPoint( const vector<BSONObj> * possibleSplitPoints = 0 ) const;
         ChunkPtr split();
 
+        /**
+         * @param splitPoints - vector to be filled in
+         * @param chunkSize  - chunk size to target in bytes
+         */
         void pickSplitVector( vector<BSONObj>& splitPoints , int chunkSize ) const;
         ChunkPtr multiSplit( const vector<BSONObj>& splitPoints );
 
diff --git a/s/d_split.cpp b/s/d_split.cpp
index 41c05c8d622..61238ad7650 100644
--- a/s/d_split.cpp
+++ b/s/d_split.cpp
@@ -134,15 +134,26 @@ namespace mongo {
                 errmsg = "either provide both min and max or leave both empty";
                 return false;
             }
-
+            
             long long maxChunkSize = 0;
-            BSONElement maxSizeElem = jsobj[ "maxChunkSize" ];
-            if ( maxSizeElem.eoo() ){
-                errmsg = "need to specify the desired max chunk size";
-                return false;
+            {
+                BSONElement maxSizeElem = jsobj[ "maxChunkSize" ];
+                if ( maxSizeElem.isNumber() ){
+                    maxChunkSize = maxSizeElem.numberLong() * 1<<20; 
+                }
+                else {
+                    maxSizeElem = jsobj["maxChunkSizeBytes"];
+                    if ( maxSizeElem.isNumber() ){
+                        maxChunkSize = maxSizeElem.numberLong();
+                    }
+                }
+                
+                if ( maxChunkSize <= 0 ){
+                    errmsg = "need to specify the desired max chunk size (maxChunkSize or maxChunkSizeBytes)";
+                    return false;
+                }
             }
-            maxChunkSize = maxSizeElem.numberLong() * 1<<20; 
-
+            
             Client::Context ctx( ns );
             NamespaceDetails *d = nsdetails( ns );