List:General Discussion« Previous MessageNext Message »
From:Zardosht Kasheff Date:March 3 2010 1:58pm
Subject:Re: Re bug#45458
View as plain text  
Hello Jonas,

Thank you for filing this feature request. Are there plans to add
support for clustered indexes in MySQL soon? This is something I have
been researching on and off for a while now. Here are my thoughts.

It seems that there are two parts to this feature request:
1) a new flag that allows the storage engine to report that an index
is clustered
2) changes to the optimizer to properly support clustered keys.

I like #1. The way that I dealt with it was not as good. I added
handler::supports_clustered_keys(), and used that function and
HA_CLUSTERING from my patch to determine if an index is clustered.
Your method is better.

As for #2, I do not think it is enough. Here are two other locations
of code I know that will need to be modified:
1) find_shortest_key in sql/ (This will be an addition
to MySQL bug #39653)
2) get_best_ror_intersect in sql/ This is for
index_merge. A patch of what I have done is in the attached file
9-index_merge_clustering.txt. This patch was the result of a long
thread on the internals alias (which you may want to CC for this
discussion). The link to the thread is

There may be more places that need to be modified. I think the
approach to finding out if other places need to be modified is to
pattern match off of how the optimizer deals with clustered v.
non-clustered primary keys. It does so by having a function
handler::primary_key_is_clustered. I think one needs to search the
optimizer for all instances of this function, see why it is being
called, and see if it applies to clustered v. non-clustered secondary
keys as well.


On Wed, Mar 3, 2010 at 5:57 AM, Jonas Oreland <jonas@stripped> wrote:
> Hi,
> I just filed
> which is very related to your bug#45458.
> If you would care to look at it and provide feedback,
> I would appreciate it.
> /Jonas

Index: sql/
--- sql/	(revision 17423)
+++ sql/	(revision 17424)
@@ -4525,6 +4525,7 @@ TRP_ROR_INTERSECT *get_best_ror_intersect(const PA
   ROR_SCAN_INFO *cpk_scan= NULL;
   uint cpk_no;
   bool cpk_scan_used= FALSE;
+  bool supports_clustered_keys = param->table->file->supports_clustered_keys();
   if (!(tree->ror_scans= (ROR_SCAN_INFO**)alloc_root(param->mem_root,
@@ -4536,8 +4537,20 @@ TRP_ROR_INTERSECT *get_best_ror_intersect(const PA
   for (idx= 0, cur_ror_scan= tree->ror_scans; idx < param->keys; idx++)
     ROR_SCAN_INFO *scan;
+    uint keyno= param->real_keynr[idx];
     if (!tree->ror_scans_map.is_set(idx))
+    {
+    }
+    /*
+      Ignore clustering keys.
+    */
+    if (keyno != cpk_no && param->table->key_info[keyno].flags & HA_CLUSTERING && supports_clustered_keys)
+    {
+      tree->n_ror_scans--;
+      continue;
+    }
     if (!(scan= make_ror_scan(param, idx, tree->keys[idx])))
       return NULL;
     if (param->real_keynr[idx] == cpk_no)
Re bug#45458Jonas Oreland3 Mar
  • Re: Re bug#45458Zardosht Kasheff3 Mar
    • Re: Re bug#45458Johan De Meersman3 Mar
      • RE: Re bug#45458Gavin Towey3 Mar
        • Re: Re bug#45458Zardosht Kasheff3 Mar