GSoC 2019: Implementing Merge Join:

• We have a working simple merge joiner implementation in a gerrit change here 
(https://asterix-gerrit.ics.uci.edu/#/c/3478/).
• This next week we’ll be working on handling limited memory by spilling with a 
run file.
• The next step will be to setup the hint and optimizer to use the merge joiner 
when appropriate; currently we have added a condition in JoinUtils that forces 
it to always use the merge joiner.


Multi Partitioning / Partial Broadcast Operator:

• We’ve also been working on a separate lower priority change that is mainly 
focused on supporting interval join partitioning of Project, Split, and 
Replicate partitioning schemes.
• That change is also on gerrit here 
(https://asterix-gerrit.ics.uci.edu/#/c/3223/) with more details in the commit 
message.


- Stephen Ermshar
On Jul 2, 2019, 1:05 PM -0700, Stephen Ermshar <[email protected]>, 
wrote:
> Here the first update for GSoC 2019: Implementing Merge Join. I plan on 
> writing updates like this every two weeks. Let me know if there are any 
> thoughts on the project or these updates. Thanks!
>
> Our current status:
>
> • My fork of AsterixDB for GSoC is here 
> (https://github.com/stephenermshar/asterixdb).
> • When we started we worked on pulling in old merge-join code to start off 
> with, but then decided to put end to end tests for the query and the 
> optimizer in place first.
>     • The optimizer test is currently blank. The runtime test in 
> `merge-join/equi-join` is based on the `btree-index-nested-loop-join` test 
> and has a merge-join hint.
>     • Those tests are in my fork on the branch 
> [stephenermshar/merge-join/Import-original-hyracks-merge-join](https://github.com/stephenermshar/asterixdb/tree/stephenermshar/merge-join/import-original-hyracks-merge-join).
>
> The general plan at this point is to add a Merge Join Operator. It will have 
> an activity to take two data input streams, and a joiner activity that can 
> request data from the input activity when it’s ready. Our first 
> implementation won’t include spilling to disk.
>
> - Stephen Ermshar
> On May 28, 2019, 2:35 PM -0700, Mike Carey <[email protected]>, wrote:
> > Cool!  I would be happy to be looped in occasionally as well - so I will
> > watch this space for the public progress updates.  In addition to adding
> > the algorithm, we'll need to add an optimizer rule (actually update an
> > existing rule) so that it gets picked when appropriate - and also a hint
> > so that you can manually suggest that it be picked regardless.  DB2
> > chooses this join method when the incoming data arguments are already
> > sorted, so that would likely be a good heuristic for us too.  (It is
> > sure to be cheaper than anything else in that case.)  (So think about
> > this as adding a Merge Join, not a Sort Merge Join, actually, as is not
> > likely to be the case that sorting and then doing this will be a cost win.)
> >
> > On 5/26/19 9:22 PM, Stephen Ermshar (gmail) wrote:
> > > Hi Ali,
> > >
> > > Since the coding period starts tomorrow I wanted to get in touch with you 
> > > again.
> > >
> > > Preston and I were thinking it would be good to post weekly reports here 
> > > on the dev mailing list to keep a public record of our progress. We’d 
> > > also like to have weekly or twice-weekly meetings to track our progress 
> > > and direction. If you have any thoughts on how we can organize our 
> > > efforts more this summer let me know.
> > >
> > > This week Preston and I decided we’d focus on refreshing my knowledge of 
> > > the Sort Merge Join algorithm. I’d like to also get a fresh development 
> > > environment setup for this summer and if possible start moving old code 
> > > in from the outdated repository.
> > >
> > > I’m excited to start working on this!
> > >
> > > - Stephen Ermshar
> > >

Reply via email to