Speed up SQL aggregate functions percentile() and median() by avoiding a full sort...

FossilOrigin-Name: 9232b5f21d76dc702ccda5c689d61bcd0057fb1f2a4db6a70c607c25ad03a027

diff --git a/manifest b/manifest
index daed8e5048dc2eb6d3de911520cb4fd0f8eb4c36..bb9a1c55f8e0a9ae1a9980603858f0b47d092287 100644 (file)
--- a/manifest
+++ b/manifest
@@ -1,5 +1,5 @@
-C When\sreading\sa\ssuper-journal\sname\sfrom\sa\sjournal\sfile,\sallocate\sa\snew\sbuffer\srather\sthan\susing\sPager.pTmpSpace.\sThis\sprevents\sa\sbuffer\soverrun\sthat\scould\soccur\swhen\susing\sa\sVFS\swith\sa\slarge\ssqlite3_vfs.mxPathname\svalue\swith\sa\sdatabase\swith\sa\ssmall\spage\ssize.
-D 2026-06-12T18:35:24.464
+C Speed\sup\sSQL\saggregate\sfunctions\spercentile()\sand\smedian()\sby\savoiding\sa\sfull\ssort\sof\sthe\sarray\sof\sarguments.
+D 2026-06-13T20:16:54.656
 F .fossil-settings/binary-glob 61195414528fb3ea9693577e1980230d78a1f8b0a54c78cf1b9b24d0a409ed6a x
 F .fossil-settings/empty-dirs dbb81e8fc0401ac46a1491ab34a7f2c7c0452f2f06b54ebb845d024ca8283ef1
 F .fossil-settings/ignore-glob 35175cdfcf539b2318cb04a9901442804be81cd677d8b889fcc9149c21f239ea
@@ -692,7 +692,7 @@ F src/delete.c 59eeca3fb88c29329afc41bb803ee568b120d9dd7470b5f38ab55cc38390b451
 F src/expr.c e97dd9f6ada4c448764e225d8963091bf630b3efb2c92e4d0762571cca2a14e5
 F src/fault.c 460f3e55994363812d9d60844b2a6de88826e007
 F src/fkey.c 931f74cec1dc8038a0217ef340c91ce147dd1bbed08dc40c47ee0ec6edfffb08
-F src/func.c 555d5f7686f9eef20fe3574889403f307a6cb16ada5b05cbc6a9288dcf75aaeb
+F src/func.c 038e454176de1729cc2a1fccce5629d86cc3fa6860546ba54827b394b1a5b4ed
 F src/global.c a19e4b1ca1335f560e9560e590fc13081e21f670643367f99cb9e8f9dc7d615b
 F src/hash.c 03c8c0f4be9e8bcb6de65aa26d34a61d48a9430747084a69f9469fbb00ea52ca
 F src/hash.h 46b92795a95bfefb210f52f0c316e9d7cdbcdd7e7fcfb0d8be796d3a5767cddf
@@ -2209,9 +2209,11 @@ F tool/warnings-clang.sh bbf6a1e685e534c92ec2bfba5b1745f34fb6f0bc2a362850723a9ee
 F tool/warnings.sh a554d13f6e5cf3760f041b87939e3d616ec6961859c3245e8ef701d1eafc2ca2
 F tool/win/sqlite.vsix deb315d026cc8400325c5863eef847784a219a2f
 F tool/winmain.c 00c8fb88e365c9017db14c73d3c78af62194d9644feaf60e220ab0f411f3604c
-P f15d076820f4a0d9e7fa32d99575bce27bde5e987bb415e6cfb69b06f0da7e1e ac17669e840dbac5c7c75e9b6014cca2d289aba8de36f9d1d0d61186e8cc6207
-R b4784016cdc66719536104e64f4ddb51
-T +closed ac17669e840dbac5c7c75e9b6014cca2d289aba8de36f9d1d0d61186e8cc6207
+P 7f71859841af7cb0806f58e9c8013a990fcca72b807a0513156d7127ce5c7b62
+R 440eb116549bb7b7e06b073447b72a9b
+T *branch * pecentile-partial-sort
+T *sym-pecentile-partial-sort *
+T -sym-trunk *
 U dan
-Z ae30c5c61ab231f09fb1164a547a0499
+Z 2032cb49443943956dfcd7d4c638a43c
 # Remove this line to create a well-formed Fossil manifest.

diff --git a/manifest.tags b/manifest.tags
index bec971799ff1b8ee641c166c7aeb22d12c785393..3847f152e1bd0577be607a9f69b46aaf4049a1ee 100644 (file)
--- a/manifest.tags
+++ b/manifest.tags
@@ -1,2 +1,2 @@
-branch trunk
-tag trunk
+branch pecentile-partial-sort
+tag pecentile-partial-sort

diff --git a/manifest.uuid b/manifest.uuid
index e7fdd067d6009953a4f29789b328fe47131613b4..1ace09b978d18a74e006abb6d7604186c4f4668f 100644 (file)
--- a/manifest.uuid
+++ b/manifest.uuid
@@ -1 +1 @@
-7f71859841af7cb0806f58e9c8013a990fcca72b807a0513156d7127ce5c7b62
+9232b5f21d76dc702ccda5c689d61bcd0057fb1f2a4db6a70c607c25ad03a027

diff --git a/src/func.c b/src/func.c
index f5935fdb44188ccf6f7e38a039a42dfeb0f415a7..6439451fabbc0364de970f5dbbd409c11f64cd2a 100644 (file)
--- a/src/func.c
+++ b/src/func.c
@@ -2967,8 +2967,17 @@ static void percentStep(sqlite3_context *pCtx, int argc, sqlite3_value **argv){
 **    (1)  To avoid a dependency on qsort()
 **    (2)  To avoid the function call to the comparison routine for each
 **         comparison.
+**
+** If parameter iReq is non-negative, then the caller will only access
+** elements a[iReq] and a[iReq+1] (if it exists) of the sorted array and
+** so it is not necessary to position any other elements. Or if iReq is
+** negative, then the final array must be fully sorted.
 */
-static void percentSort(double *a, unsigned int n){
+static void percentSort(
+  double *a,                      /* Array to sort */
+  unsigned int n,                 /* Number of elements in array a[] */
+  int iReq                        /* Element caller cares about (or -ve) */
+){
   int iLt;  /* Entries before a[iLt] are less than rPivot */
   int iGt;  /* Entries at or after a[iGt] are greater than rPivot */
   int i;         /* Loop counter */
@@ -3005,17 +3014,36 @@ static void percentSort(double *a, unsigned int n){
       }
     }while( i<iGt );
 
-    /* Recurse on the smaller partition only.  The smaller partition
-    ** will hold n/2 or fewer entries, which assures that the stack
-    ** depth will not exceed O(log(n)), even for pathological cases.
-    ** Loop without recursion for the larger partition. */
-    if( iLt>(int)(n/2) ){
-      if( n-iGt>=2 ) percentSort(a+iGt, n-iGt);
-      n = iLt;
+    assert( a[iLt]==rPivot );
+    assert( iGt>iLt );
+
+    if( iReq>=0 ){
+      /* In this case, the only elements that the caller requires sorted into 
+      ** the correct positions are elements a[iReq] and a[iReq+1]. At this
+      ** point we know that element a[iLt] is in the correct position, so if
+      ** iReq is less than iLt, it is only necessary to sort the first
+      ** partition. Or, if iReq is greater than or equal to iLt, it is only
+      ** necessary to sort the second.  */
+      if( iReq<iLt ){
+        n = iLt;
+      }else{
+        a += iGt;
+        n -= iGt;
+        iReq = MAX(0, iReq-iGt);
+      }
     }else{
-      if( iLt>=2 ) percentSort(a, iLt);
-      a += iGt;
-      n -= iGt;
+      /* Recurse on the smaller partition only.  The smaller partition
+      ** will hold n/2 or fewer entries, which assures that the stack
+      ** depth will not exceed O(log(n)), even for pathological cases.
+      ** Loop without recursion for the larger partition. */
+      if( iLt>(int)(n/2) ){
+        if( n-iGt>=2 ) percentSort(a+iGt, n-iGt, -1);
+        n = iLt;
+      }else{
+        if( iLt>=2 ) percentSort(a, iLt, -1);
+        a += iGt;
+        n -= iGt;
+      }
     }
   }while( n>=2 );
 }
@@ -3052,7 +3080,7 @@ static void percentInverse(sqlite3_context *pCtx,int argc,sqlite3_value **argv){
   }
   if( p->bSorted==0 ){
     assert( p->nUsed>1 );
-    percentSort(p->a, p->nUsed);
+    percentSort(p->a, p->nUsed, -1);
     p->bSorted = 1;
   }
   p->bKeepSorted = 1;
@@ -3081,13 +3109,13 @@ static void percentCompute(sqlite3_context *pCtx, int bIsFinal){
   if( p==0 ) return;
   if( p->a==0 ) return;
   if( p->nUsed ){
+    ix = p->rPct*(p->nUsed-1);
+    i1 = (unsigned)ix;
     if( p->bSorted==0 ){
       assert( p->nUsed>1 );
-      percentSort(p->a, p->nUsed);
+      percentSort(p->a, p->nUsed, (bIsFinal ? (int)i1 : -1));
       p->bSorted = 1;
     }
-    ix = p->rPct*(p->nUsed-1);
-    i1 = (unsigned)ix;
     if( settings & 1 ){
       vx = p->a[i1];
     }else{