Kenneth, I just pushed the revised patch (v2!). The revised approach samples the parent relation to estimate the number of tuples rather than performing a complete scan. In my tests, the estimate appears to be accurate, erring on the larger side, which is fine.


That is great. I am looking forward to your patch. After the
issues that you needed to address, I think that it would be
reasonable to add a few more user settings for the hash index.
Fill-factor is too course a knob. The others that I have been
considering are:

maxtuples - Not really the maximum, but a target value to use
  for setting up the initial buckets. This would allow you to
  set it for data loads and avoid the "split-n-copy" trauma
  that you are trying to avoid with your new hash build process.
If I understand you correctly, I believe we already do this with our current build process, there should not be any splits of the index if we estimated the tuple count correctly. However, what gets you is collisions where lots of overflow pages occur when distinct keys map to the same bucket, or if you have lots of duplicate keys. Because your overall tuple count doesn't exceed the fill factor, no splits occur, but lengthy bucket chains lead to lots of IOs. You touch on this below.
multiplicity - Try to capture use cases that would require many
  overflow pages. In particular, if we discard the normal index
  page layout we can skip the space overhead of the page pointer
  and generate a more compact index. Then you could use a few
  more hash bits to lookup the index entry in the page. How many
  bits would be determined by this factor. 8-bits would give
  you 256 sub-pieces that could each hold about 3 entries using
  the current 4-byte hash, or 2 entries using an 8-byte hash.

What do you think?
Yes, this is a good direction. If we can increase the number of buckets and reduce the bucket size (either physically or virtually) to allow more direct access without creating a huge index on disk, that would be ideal. But, then if you do have collisions, overflows occur more frequently. I spoke with Neil Conway yesterday at the PG conference here in Portland and he piqued my interest in examining his hash code more closely to see what he has already done in this area.


On Wed, Oct 17, 2007 at 03:31:58PM -0700, Tom Raney wrote:


Yes, we did update the code to use the estimate. I will post the patch with this update. We only saw a very small difference in index build time, but you may when you add many columns to the base relation. With 1 billion tuples, you should start to see the hash index outperform the btree for some equality probes, I would imagine. With a 90% fill factor, the btree would require 4 levels to index that many tuples. If the top two were in memory, there would be 3 IOs needed. I don't think PG supports index only scans, so it will take the extra IO to probe the base relation. The hash may take up to 2 IOs and maybe even less (or maybe more depending on how many overflow buckets there are). It might be interesting to fiddle with the fill factors of each index - hash pages (buckets) default to 75% full. -Tom

I am getting ready to stitch in an updated, simplified version
of Neil Conway's hash-only hash index patch. Did you have a
chance to update your sizing function to use the planner-like
estimate and not a full table scan? I would like to be able
to use that when my test table start to have 10^9 entries.
If you have not had a chance, I will try and add it myself.


---------------------------(end of broadcast)---------------------------
TIP 2: Don't 'kill -9' the postmaster

---------------------------(end of broadcast)---------------------------
TIP 5: don't forget to increase your free space map settings

Reply via email to