Revise the strategy used by the star-query heuristic: Instead of decreasing

the cost of all fact-table WhereLoops, increase the run-cost of WhereLoops that
are SCANs of dimension tables.

FossilOrigin-Name: 1bc09c9e8bd77ac41ecbe510c7e003757fc11d0f586da6cdf3584315aa9d6407

diff --git a/manifest b/manifest
index b407f7242ea77bc8ab72b42240baaf0e9e5d7e2a..c6113e84a4449a824622880582ce0f6ce298c3a0 100644 (file)
--- a/manifest
+++ b/manifest
@@ -1,5 +1,5 @@
-C Improvments\sto\sdebug\soutput\son\sthe\sstar-query\sheuristic.
-D 2025-01-25T14:30:36.057
+C Revise\sthe\sstrategy\sused\sby\sthe\sstar-query\sheuristic:\s\sInstead\sof\sdecreasing\nthe\scost\sof\sall\sfact-table\sWhereLoops,\sincrease\sthe\srun-cost\sof\sWhereLoops\sthat\nare\sSCANs\sof\sdimension\stables.
+D 2025-01-25T23:04:05.646
 F .fossil-settings/empty-dirs dbb81e8fc0401ac46a1491ab34a7f2c7c0452f2f06b54ebb845d024ca8283ef1
 F .fossil-settings/ignore-glob 35175cdfcf539b2318cb04a9901442804be81cd677d8b889fcc9149c21f239ea
 F LICENSE.md e108e1e69ae8e8a59e93c455654b8ac9356a11720d3345df2a4743e9590fb20d
@@ -863,8 +863,8 @@ F src/vxworks.h d2988f4e5a61a4dfe82c6524dd3d6e4f2ce3cdb9
 F src/wal.c 4e6181d8780ab0af2e1388d0754cbe6f2f04593d2b1ab6c41699a89942fd8997
 F src/wal.h ba252daaa94f889f4b2c17c027e823d9be47ce39da1d3799886bbd51f0490452
 F src/walker.c d5006d6b005e4ea7302ad390957a8d41ed83faa177e412f89bc5600a7462a014
-F src/where.c 8fb5ce22e5fbc82f8e53a059e76ff492113f3de8ff26c3639d52c75e94088596
-F src/whereInt.h 3b2ef4617758174d00c6940850e4988c1b195d39bcc20f2965319a1f7bffc714
+F src/where.c 3917a56c85951be386057ee4e9af8a6746427f6f26e8ffce293a61e334b76498
+F src/whereInt.h 510b12f85a763602a1b1e70a9cc643a73082a63cda2ca204dd912daaa540115d
 F src/wherecode.c 0c3d3199a2b769a5e2bb70feb5003dc85b3d86842ecaf903a47f2b4205ca5dab
 F src/whereexpr.c 0f93a29cabd3a338d09a1f5c6770620a1ac51ec1157f3229502a7e7767c60b6f
 F src/window.c 2bf01f9941a64fbcead61a0e3cb5db3fca5094b30d2ff0d23274c2a81d2e2385
@@ -2208,8 +2208,8 @@ F tool/version-info.c 3b36468a90faf1bbd59c65fd0eb66522d9f941eedd364fabccd7227350
 F tool/warnings-clang.sh bbf6a1e685e534c92ec2bfba5b1745f34fb6f0bc2a362850723a9ee87c1b31a7
 F tool/warnings.sh 49a486c5069de041aedcbde4de178293e0463ae9918ecad7539eedf0ec77a139
 F tool/win/sqlite.vsix deb315d026cc8400325c5863eef847784a219a2f
-P 8b9e621dbf599d7e75e07d75d4c400247d693ea76e00eba0919f4e33e3e10bdd
-R e6dd9f24405ca06acecad76b774b4e57
+P b3ebeb0682a2c837987acf4ed92f06cf91aea235830c5a0f9dd1ce64afe16e84
+R a3d46d3d0d804ef5b7f52b211e4f3855
 U drh
-Z 5c57dd71c02ea0fa33cecfe1d8ec4eed
+Z 01bfb692a509c7bcf6c556a6b5d760a5
 # Remove this line to create a well-formed Fossil manifest.

diff --git a/manifest.uuid b/manifest.uuid
index a25e3e3f1408245236dbc8b8e9f90e9430bdfef5..29b013c9fc75f9577b07e178e57c45266006a3f5 100644 (file)
--- a/manifest.uuid
+++ b/manifest.uuid
@@ -1 +1 @@
-b3ebeb0682a2c837987acf4ed92f06cf91aea235830c5a0f9dd1ce64afe16e84
+1bc09c9e8bd77ac41ecbe510c7e003757fc11d0f586da6cdf3584315aa9d6407

diff --git a/src/where.c b/src/where.c
index d25a1723d3982abfc78c7472f09a903c5ca33f6d..0db7cbc1e2390328772de11a5e8d349d76998e84 100644 (file)
--- a/src/where.c
+++ b/src/where.c
@@ -2475,9 +2475,9 @@ void sqlite3WhereLoopPrint(const WhereLoop *p, const WhereClause *pWC){
   }else{
     sqlite3DebugPrintf(" f %06x N %d", p->wsFlags, p->nLTerm);
   }
-  if( pWInfo && pWInfo->nOutStarDelta>0 && p->rStarDelta!=0 ){
+  if( pWInfo && pWInfo->bStarUsed && p->rStarDelta!=0 ){
     sqlite3DebugPrintf(" cost %d,%d,%d delta=%d\n",
-                       p->rSetup, p->rRun, p->nOut, -p->rStarDelta);
+                       p->rSetup, p->rRun, p->nOut, p->rStarDelta);
   }else{
     sqlite3DebugPrintf(" cost %d,%d,%d\n", p->rSetup, p->rRun, p->nOut);
   }
@@ -5458,13 +5458,13 @@ static LogEst whereSortingCost(
 **
 ** SIDE EFFECT:  (and really the whole point of this subroutine)
 **
-** If pWInfo describes a star-query, then the cost on WhereLoops for the
-** fact table is reduced.  This heuristic helps keep fact tables in
+** If pWInfo describes a star-query, then the cost for SCANs of dimension
+** WhereLoops is increased to be slightly larger than the cost of a SCAN
+** in the fact table.  This heuristic helps keep fact tables in
 ** outer loops.  Without this heuristic, paths with fact tables in outer
 ** loops tend to get pruned by the mxChoice limit on the number of paths,
-** resulting in poor query plans.  The total amount of heuristic cost
-** adjustment is stored in pWInfo->nOutStarDelta and the cost adjustment
-** for each WhereLoop is stored in its rStarDelta field.
+** resulting in poor query plans.  The cost adjustment for each WhereLoop
+** is stored in its rStarDelta field.
 **
 ** This heuristic can be completely disabled, so that no query is
 ** considered a star-query, using SQLITE_TESTCTRL_OPTIMIZATION to
@@ -5488,7 +5488,7 @@ static int computeMxChoice(WhereInfo *pWInfo){
   }
 #endif /* SQLITE_DEBUG */
 
-  if( nLoop>=5 
+  if( nLoop>=5
    && !pWInfo->bStarDone
    && OptimizationEnabled(pWInfo->pParse->db, SQLITE_StarQuery)
   ){
@@ -5504,12 +5504,12 @@ static int computeMxChoice(WhereInfo *pWInfo){
     ** the cost of fact tables relative to dimension tables, as a heuristic
     ** to help keep the fact tables in outer loops.
     */
-    assert( pWInfo->nOutStarDelta==0 );
+    assert( !pWInfo->bStarUsed );
     aFromTabs = pWInfo->pTabList->a;
     pStart = pWInfo->pLoops;
     for(iFromIdx=0, m=1; iFromIdx<nLoop; iFromIdx++, m<<=1){
       int nDep = 0;             /* Number of dimension tables */
-      LogEst rDelta;            /* Heuristic cost adjustment */
+      LogEst mxRun;             /* Maximum SCAN cost of a fact table */
       Bitmask mSeen = 0;        /* Mask of dimension tables */
       SrcItem *pFactTab;        /* The candidate fact table */
 
@@ -5543,43 +5543,47 @@ static int computeMxChoice(WhereInfo *pWInfo){
         }
       }
       if( nDep<=3 ) continue;
-      rDelta = 15*(nDep-3);
-#ifdef WHERETRACE_ENABLED /* 0x80000 */
-      if( sqlite3WhereTrace & 0x80000 ){
-        Bitmask x;
-        int ii;
-        sqlite3DebugPrintf(
-          "Fact-table %s(%d): cost reduced %d due to %d dimension tables:",
-          pFactTab->zAlias ? pFactTab->zAlias : pFactTab->pSTab->zName,
-          iFromIdx, rDelta, nDep
-        );
-        for(ii=0, x=1; ii<nLoop; ii++, x<<=1){
-          if( x & mSeen ){
-            SrcItem *pDim = aFromTabs + ii;
-            sqlite3DebugPrintf(" %s(%d)",
-              pDim->zAlias ? pDim->zAlias : pDim->pSTab->zName, ii
-            );
-          }
-        }
-        sqlite3DebugPrintf("\n");
+
+      /* Compute the maximum cost of any WhereLoop for the iFromIdx-th term */
+      mxRun = -32767;
+      for(pWLoop=pStart; pWLoop; pWLoop=pWLoop->pNextLoop){
+        if( pWLoop->iTab<iFromIdx ) continue;
+        if( pWLoop->iTab>iFromIdx ) break;
+        if( pWLoop->rRun>mxRun ) mxRun = pWLoop->rRun;
       }
-#endif
-      if( pWInfo->nOutStarDelta==0 ){
+
+      /* Make sure rStarDelta values are initialized */
+      if( !pWInfo->bStarUsed ){
         for(pWLoop=pWInfo->pLoops; pWLoop; pWLoop=pWLoop->pNextLoop){
           pWLoop->rStarDelta = 0;
         }
+        pWInfo->bStarUsed = 1;
       }
-      pWInfo->nOutStarDelta += rDelta;
-      for(pWLoop=pWInfo->pLoops; pWLoop; pWLoop=pWLoop->pNextLoop){
-        if( pWLoop->maskSelf==m ){
-          pWLoop->rRun -= rDelta;
-          pWLoop->nOut -= rDelta;
-          pWLoop->rStarDelta = rDelta;
+
+      /* Increase the cost of table scans for dimension tables to be slightly
+      ** more than the maximum cost of fact table */
+      for(pWLoop=pStart; pWLoop; pWLoop=pWLoop->pNextLoop){
+        if( (pWLoop->maskSelf & mSeen)==0 ) continue;
+        if( pWLoop->nLTerm ) continue;
+        if( pWLoop->rRun<=mxRun ){
+#ifdef WHERETRACE_ENABLED /* 0x80000 */
+          if( sqlite3WhereTrace & 0x80000 ){
+            SrcItem *pDim = aFromTabs + pWLoop->iTab;
+            sqlite3DebugPrintf(
+              "Increase SCAN cost of dimension %s(%d) of fact %s(%d) to %d\n",
+              pDim->zAlias ? pDim->zAlias: pDim->pSTab->zName, pWLoop->iTab,
+              pFactTab->zAlias ? pFactTab->zAlias : pFactTab->pSTab->zName,
+              iFromIdx, mxRun+1
+            );
+          }
+#endif /* WHERETRACE_ENABLED */
+          pWLoop->rStarDelta = mxRun+1 - pWLoop->rRun;
+          pWLoop->rRun = mxRun+1;
         }
       }
     }
 #ifdef WHERETRACE_ENABLED /* 0x80000 */
-    if( (sqlite3WhereTrace & 0x80000)!=0 && pWInfo->nOutStarDelta ){
+    if( (sqlite3WhereTrace & 0x80000)!=0 && pWInfo->bStarUsed ){
       sqlite3DebugPrintf("WhereLoops changed by star-query heuristic:\n");
       for(pWLoop=pWInfo->pLoops; pWLoop; pWLoop=pWLoop->pNextLoop){
         if( pWLoop->rStarDelta ){
@@ -5589,7 +5593,7 @@ static int computeMxChoice(WhereInfo *pWInfo){
     }
 #endif
   }
-  return pWInfo->nOutStarDelta>0 ? 18 : 12;
+  return pWInfo->bStarUsed ? 18 : 12;
 }
 
 /*
@@ -6035,9 +6039,9 @@ static int wherePathSolver(WhereInfo *pWInfo, LogEst nRowEst){
     }
   }
 
-  pWInfo->nRowOut = pFrom->nRow + pWInfo->nOutStarDelta;
+  pWInfo->nRowOut = pFrom->nRow;
 #ifdef WHERETRACE_ENABLED
-  pWInfo->rTotalCost = pFrom->rCost + pWInfo->nOutStarDelta;
+  pWInfo->rTotalCost = pFrom->rCost;
 #endif
 
   /* Free temporary memory and return success */
@@ -6436,7 +6440,6 @@ static SQLITE_NOINLINE void whereCheckIfBloomFilterIsUseful(
       }
     }
     nSearch += pLoop->nOut;
-    if( pWInfo->nOutStarDelta ) nSearch += pLoop->rStarDelta;
   }
 }
 

diff --git a/src/whereInt.h b/src/whereInt.h
index 5bea70ba9fa3374d0790cbdef7591f81242b07ab..35bfe581dde5a922e78b7c570ddff730d917a0e9 100644 (file)
--- a/src/whereInt.h
+++ b/src/whereInt.h
@@ -163,7 +163,7 @@ struct WhereLoop {
 # define WHERE_LOOP_XFER_SZ offsetof(WhereLoop,nLSlot)
   u16 nLSlot;           /* Number of slots allocated for aLTerm[] */
   LogEst rStarDelta;    /* Cost delta due to star-schema heuristic.  Not
-                        ** initialized unless pWInfo->nOutStarDelta>0 */
+                        ** initialized unless pWInfo->bStarUsed */
   WhereTerm **aLTerm;   /* WhereTerms used */
   WhereLoop *pNextLoop; /* Next WhereLoop object in the WhereClause */
   WhereTerm *aLTermSpace[3];  /* Initial aLTerm[] space */
@@ -487,7 +487,7 @@ struct WhereInfo {
   unsigned bOrderedInnerLoop:1;/* True if only the inner-most loop is ordered */
   unsigned sorted :1;          /* True if really sorted (not just grouped) */
   unsigned bStarDone     :1;   /* True if check for star-query is complete */
-  LogEst nOutStarDelta;     /* Artifical nOut reduction for star-query */
+  unsigned bStarUsed     :1;   /* True if cost adjustments for star-query */
   LogEst nRowOut;           /* Estimated number of output rows */
 #ifdef WHERETRACE_ENABLED
   LogEst rTotalCost;        /* Total cost of the solution */