Apache
CouchDB
Monday, November 2, 2009
how couchdb treats the disks
append only btree
Replication
JSON
Map Reduce
HTTP
Monday, November 2, 2009
intro:
offline apps
simple development
easy to deploy
Stuart Langridge - Canonical
! !
Monday, November 2, 2009
15 million desktops by the end of the year.
solve the data island problem
Monday, November 2, 2009
raindrop - next gen messaging platform
name: your own personal piece of the cloud
Monday, November 2, 2009
Monday, November 2, 2009
Couchapps
-- easiest way to deploy an Ajax application that needs to store data.
-- capable of scaling up with CouchDB. (etags etc)
“Ground Computing”
@jhuggins
http://www.flickr.com/photos/mcpig/872293700/
Monday, November 2, 2009
app runs near the user. lower latency, lower cost.
Monday, November 2, 2009
Offline by Default
http://www.flickr.com/photos/shane-h/280084650
Monday, November 2, 2009
reliable
telephone poles / like akamai but for dynamic database backed applications.
- same application on the client and server or the cloud
http://apod.nasa.gov/apod/ap050930.html
“Of the Web”
http://jacobian.org/writing/of-the-web/
Monday, November 2, 2009
Let me tell you something: Django may be built for the Web, but CouchDB is built of the Web. I've never seen software that so
completely embraces the philosophies behind HTTP. ... this is what the software of the future looks like. Jacob Kaplan-Moss --
October, 2007
perfect spot - discovered not invented
When you can move data via replication, it
changes the physics of the web
so data is useful independent of location.
Lockless
Storage
Monday, November 2, 2009
html5 story
js
ruby
transactions?
-- getting CouchDB right involves going all the way to core lockless storage engine.
performance, map reduce, and deployment
B-Tree
Monday, November 2, 2009
data structure
sorted by keys
minimize disk seeks
Monday, November 2, 2009
logical btree
Monday, November 2, 2009
indexes - efficient
key lookups
range scans
Append
Only
Monday, November 2, 2009
robust / reliable
never touch bytes that are on disk
can’t be inconsistent
Monday, November 2, 2009
append only - in action
update J to a new value (rewrite j’s leaf node)
Monday, November 2, 2009
rewrite the path along the tree to J
up the parent path until we write the root.
Monday, November 2, 2009
all pointers point backwards into the file
Monday, November 2, 2009
update cost?
only has O(log n) cost where n is the # of keys
Monday, November 2, 2009
another update - add 5.
rewrite path to Z node to add 5
Monday, November 2, 2009
time follows the blue arrow.
top half is actually an accumulation of updates like the bottom.
note: batch efficiency -- save block updates
Monday, November 2, 2009
2 day test run
couchdb over long haul
4 writers, 16 readers, random access, small keyspace - lots of 409s etc.
Incremental
Reduce
Monday, November 2, 2009
takes advantage of the write pattern we’ve seen
setup main storage doc count example
Monday, November 2, 2009
rereduce = false
count # of keys in the block
Monday, November 2, 2009
rereduce = true
sum # of docs in child nodes
flexible range querying
Monday, November 2, 2009
intermediate reductions stored during a view calculation
updating J doesn’t change any counts, adding the 5 key does
Monday, November 2, 2009
we are rewriting the parent nodes anyway, so it makes sense to run the reduce calculation
Concurrent
Access
Monday, November 2, 2009
readers see a consistent view
we can support many readers
reads don’t block writes
writes don’t block reads
Monday, November 2, 2009
concurrent readers
list all docs, then a write comes in, you don’t want to be surprised by an extra doc
readers can be as long lived as they want without blocking writers or seeing inconsistencies
Local Web
Platform
Monday, November 2, 2009
user interviews
http://www.longnow.org/projects/clock/orrery/
Monday, November 2, 2009
append-only
no fixup phase
reliable and concurrent
sequential write pattern (few writer seeks)
hot backups
compaction
0 comments
Post a comment