[
https://issues.apache.org/jira/browse/COUCHDB-1270?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Paul Joseph Davis updated COUCHDB-1270:
---------------------------------------
Attachment: 0004-Remove-the-old-view-engine.patch
0003-Create-the-couch_mrview-application.patch
0002-Create-the-couch_index-application.patch
0001-Minor-changes-for-new-indexing-engine.patch
@Filipe
Good catch on both. For point one, I just removed the nil default. In my
refactoring I undid almost all of these setting record fields to nil because
its fairly silly to use a nonstandard value for undefined when the standard
value is literally, undefined. One place I didn't change this yet was for the
qserver variable because of how couch_query_servers:stop_doc_map is written and
I was wanting to minimize the scope of these changes.
As to 2, I added the file removal in the cancellation function. One thing this
doesn't address is a generalized method for handling the new compaction
lifetime model. We'll have to revisit this a bit to try and get a better
abstraction for compaction.
The timing tests I was running were against the views in [1] which you've used
before for these types of tests. The test is just a simple "time curl
http://127.0.0.1:5984/indexer_test/_design/test/_view/view1?limit=0"; or w/e the
URL is for that design doc. Fairly consistently the new-views are a minute or
two faster than trunk. So, nothing super improved but convincing enough to me
that I haven't trashed performance.
Also, I did spend a couple days profiling the view engine and comparing a
couple dirty hacks. I'm convinced we can shorten these times even more with
some more refactoring and playing with some other optimizations but they were
leading me deep into the couch_query_servers and couch_os_process. I plan on
backporting some refactoring for these parts next and that should set us up for
some of the more advanced changes to the view engine.
Trunk: 26m29.809s
New-Views: 24m50.923s
[1] http://fdmanana.couchone.com/indexer_test
@Benoit
I'd agree on turnign couch_index_api into a behavior but I decided to delay
that until after we update the various applications and build system so that we
have a proper ERL_LIBS directory. For behaviours to work they need to be
importable and that's not the easiest given our current structure. Though it'll
definitely be something to add in the future.
For supervisors I think that could be a good idea, but I'm not the most
familiar with them and their side effects to know if they'd be a good fit here.
You mentioned on IRC that you'd take a swing at implementing that and I'd be
quite interested to see what you come up witih.
> Rewrite the view engine
> -----------------------
>
> Key: COUCHDB-1270
> URL: https://issues.apache.org/jira/browse/COUCHDB-1270
> Project: CouchDB
> Issue Type: Improvement
> Components: JavaScript View Server
> Reporter: Paul Joseph Davis
> Attachments: 0001-Minor-changes-for-new-indexing-engine.patch,
> 0002-Create-the-couch_index-application.patch,
> 0003-Create-the-couch_mrview-application.patch,
> 0004-Remove-the-old-view-engine.patch
>
>
> The view engine has been creaky and cluttered. As shown by GeoCouch, adding
> new indexers basically involves copying the entire view engine and hacking
> the parts that are different. In short, the opposite of good engineering.
> Over the last couple weeks I've refactored the view engine and reimplemented
> the map/reduce view engine. These changes are 100% internal and no external
> behavior has changed. Performance is just a tiny bit better than trunk. I did
> do some playing trying to improve view update times and there are some dances
> we could do, but for the time being I wanted to keep the same general
> architecture for updates so that the changes are minimal.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
