Hi Wout, thanks for writing this up.
One comment about the map-only views: I think you'll find that Couch
has already done a good bit of the work needed to support them, too.
Couch maintains a btree for each design doc keyed on docid that stores
all the view keys emitted by the maps over each document. When a
document is updated and then analyzed, Couch has to consult that
btree, purge all the KVs associated with the old version of the doc
from each view, and then insert the new KVs. So the tracking
information correlating docids and view keys is already available.
You'd still be left with the problem of generating unique docids for
the documents in the Review DB, but I think that's a problem that
needs to be solved. The restriction to only MR views with no
duplicate keys across views seems too strong to me.
With that said, I'd prefer to spend my time extending the view engine
to handle chainable MR workflows in a single shot. Especially in the
simple sort_by_value case it just seems like a cleaner way to go about
things. Cheers,
Adam
On Apr 22, 2009, at 8:40 AM, Wout Mertens wrote:
Intro
=====
How do you sort by reduce value? How do you join views? How do you
get unique view results? How do you cache group key reduces?
I think that with the below proposed solution all the above and more
are possible. The general idea is to store view results and run map/
reduce on them. There's been some discussions about this but they
went nowhere. I've been thinking about this issue a bit and I think
it can be done.
I'd like to call this feature a Review DB.
Use cases
=========
- Suppose you want to know what tags are most popular on your blog.
Simply get:
http://couchdb/db/_design/myblog/_review/tags_by_count/_view/sort_by_value
Where tags_by_count is a Review DB that gets input from the tagcount
view and then runs the sort_by_value view on it, a map() function
that simply emits (value,key).
Likewise, show pages in order of popularity, whereby user can vote
up (+1) or down (-1):
http://couchdb/db/_design/mywiki/_review/pagevotes/_view/
sort_by_value
- Given documents with attributes title, date and tags. You'd like
to know the minimum value of date and a breakdown by count for tags,
for every title. Normally you'd use 2 map+reduce views,
minimum_date_by_title and tagcount_by_title, which you would then
query separately. With a Review DB, you can let both views insert
their results in the database and then run a view that combines the
results in one view:
http://couchdb/db/_design/mybookstore/_review/mybooks/_view/aggregate_book_data
- This is not a way to run an on-the-fly map/reduce on a subset of a
view, like if you want to find the median popularity score of
restaurants with "Tony" in their name that are close to you.
Implementation
==============
A Review DB is a hidden database maintained by CouchDB with these
fields:
- _id of document is the string representation of the key
- "key" is the key of the incoming view row (unique)
- "value" is the value of the incoming view row
I hope that this is sufficiently like a normal view that it can be
stored as a normal view. _id is just there to make it doc-compliant,
it would be much better if "key" were the actual key.
A Review DB is defined in a design document like normal views. Each
review is an entry in the "reviews" hash, and has a "incoming_views"
array that lists all the views that should insert results in the
review db plus the group level, as well as a normal "views" hash for
further map/reduce of the review db (and perhaps another "reviews"
hash for further result processing?).
Maintaining a database of results means that results have to be
updated or even removed when documents change. I tried to make this
work (in theory) for map-only views, but the resulting requirements
are quite messy. You either need to cache the previous results of a
view for each document, or you have to have an old version of the
document available to regenerate those results.
Therefore, a Review DB only accepts results from one or more map
+reduce views. You define beforehand what the group_level of the
keys is that will be inserted.
Furthermore, a Review DB disallows (but doesn't enforce) having 2
views that generate the same keys. Otherwise, refcounting would need
to be used and while that's not difficult, I think there's limited
value in allowing this.
The Review DB needs updating every time the reduction for a group
key of one of the participating views gets updated. Even though a map
+reduce view has unique keys, we need a refcount since we have
multiple views. Whoever got to insert its value last wins.
There is a slight complication: group key values are calculated on-
the-fly from the view result b-tree. So whenever a reduce call
results in a new value for a b-tree node, AND that node is the upper
node of a subtree that is completely part of a group key, that group
key needs to be marked for recalculation.
Likewise, if deletion/addition of a b-tree node results in the
removal/creation of the sole upper node of a group key subtree, that
group key needs to be marked for removal/addition.
This is the algorithm:
- When a reducing view gets updated, and it is part of a Review DB,
use the 2 paragraphs above to keep a list of group keys that need
handling
- After updating the reduce() results, for each of the marked group
keys:
- If a group key gets removed:
- look up doc with key=group key in review db. If exists:
- delete doc
- If a group key gets added:
- look up doc with key=group key in review db. If exists:
- set doc.value to the row value
- else
- create doc with id=group key in string form, key=group key,
value=value
- If a group key gets updated:
- look up doc with key=group key in review db. If exists:
- set doc.value to the row value
- else
- create doc with id=group key in string form, key=group key,
value=value
As you can see, this is something CouchDB should do since it knows
when it's updating group key reduction values and it knows if this
was an delete, update or addition.
View updates are done when the view is called; Review updates are
done at this time as well. Views on Review DBs are done when they
are called.
Summary
=======
Review DBs are a sort of view index that CouchDB can maintain with
little overhead. It caches group key results and allows chained map
+reduce calculations using mostly existing frameworks.
I think this would be a very useful feature for CouchDB to have.
There are regularly requests for storing view results in a database
for post-processing on the mailing lists.
I'm not saying this is a trivial change but it doesn't seem
technically impossible to me either. (unless I missed something
again; this is the 5th iteration of this proposal. Anyway I know *I*
wouldn't be able to code this :-) )
What do you think, oh dear devs?
Wout.
