Scaling geodata w mapreduce natevw
-
Upload
couchbase -
Category
Technology
-
view
669 -
download
0
Transcript of Scaling geodata w mapreduce natevw
![Page 1: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/1.jpg)
Scaling geodatawith MapReduce
Nathan Vander Wilt
![Page 2: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/2.jpg)
Background
I’m Nate, a freelancer — I do web, native, embedded development using Cocoa, Django, node.js, C/C++, from SQLite to Couch. Of all, I love to talk about Couch the most :-)
Audience poll: who’s used Couchbase (or similar) “at scale” — tons of users?
Well, for better or for worse, the user in this user story is me. Here's how I turned almost ten years of personal geodata into something I can relive in less than two seconds. I hope you'll find some of these ideas helpful when dealing with tens of thousands of users.
![Page 3: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/3.jpg)
ON THE DOCKET:
•Views as “indexes”
•Basic location examples
•Geo Hacks
(pls to interrupt)
Gonna divide this up into three general chunks: make sure we’re all on the same page as far as concepts go, then dive into the main examples. Finally we’ll explore at some interesting twists of the available features.
Feel free to interrupt at any time. I’d like to have some discussion between each of these main sections, so you can also save questions for then.
![Page 4: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/4.jpg)
Views as “indexes”
![Page 5: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/5.jpg)
Views as “indexes”
Database
(Everything Ever)
efficient filtered lookup
efficient filtered lookupefficient filtered lookup
What do I mean by “indexes”?
Efficient lookup of data, extracted from documents. Think of a word index in a book, or the topical index of a “real” encyclopedia set.
![Page 6: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/6.jpg)
Views as “indexes”
Database Index
“map” function emits
With Couch you have full control. Your code defines which terms — “keys” — go into this index.
![Page 7: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/7.jpg)
GeoCouch for indexing
Spatial Index
Using R-trees as a 2-dimensional index to speed bounding box queries.
![Page 8: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/8.jpg)
GeoCouch for indexing
Spatial Lookup
![Page 9: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/9.jpg)
GeoCouch for indexing
!
Drawbacks if a lot of points inside bounding box...
![Page 10: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/10.jpg)
MapReduce indexing
B-tree index
Use B-trees as a index to speed 1-dimensional lookups
![Page 11: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/11.jpg)
MapReduce indexing
B-tree lookup
…
efficient “start/end” range queries
![Page 12: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/12.jpg)
MapReduce indexing
B-tree reduction
1 4 2 3 1
as well as grouped “reduce” queries!
![Page 13: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/13.jpg)
View “indexes” as...
Database IndexPROJECTION
Because Couch exposes its index keys directly, you can imagine the indexes as “projections” of data viewed from different perspectives.
![Page 14: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/14.jpg)
“Indexes” as projections
http://en.wikipedia.org/wiki/File:Axonometric_projection.svg
What do I mean by projection?
Example of 3D onto 2D: pictures.
![Page 15: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/15.jpg)
“Indexes” as projections
Image: USGS
Example of n-D onto 1-D: map functions
![Page 16: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/16.jpg)
Views as projections
n-dimensional
data
single dimension view
Example of n-D onto 1-D: map functions!Imagine each “aspect” of your data as its own dimension.
We’ll focus back on this in the last section on Geo Hacks, but it’s helpful to keep in mind
![Page 17: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/17.jpg)
PROJECTION
View “indexes” as...
Database*(*index posing as...) Database
Because they’re sort of just a different “perspective” on your data, in all the examples ahead we’ll look at Couch indexes almost as databases themselves.
![Page 18: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/18.jpg)
Questions so far?
![Page 19: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/19.jpg)
Basic location indexes
![Page 20: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/20.jpg)
Location example #1
Where was I when...?
![Page 21: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/21.jpg)
Location example #1
for each pt in doc: emit(pt.timestamp, pt.coord)
Map function
![Page 22: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/22.jpg)
Location example #1
(Demo)
![Page 23: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/23.jpg)
Location example #2
Where have I been?
![Page 24: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/24.jpg)
Location example #2
for each pt in doc: emit(pt.timestamp, pt.coord)
Map function
for each value in reduction: avg += value.coord n += value.n || 1return {avg, n}
Reduce function
![Page 25: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/25.jpg)
Location example #2
(Demo)
![Page 26: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/26.jpg)
Location example #3
What photos did I capture in this area?
The reason I’ve been recording my location for so many years.
![Page 27: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/27.jpg)
Location example #3
Image: http://msdn.microsoft.com/en-us/library/bb259689.aspx
emit([2,0,2,…])
?group_level=Z
Tiles, quadkeys
![Page 28: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/28.jpg)
Location example #3
for each pt in doc: key = quadkey(pt.coord) emit(key, pt)
Map function
_countReduce function (basic)
![Page 29: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/29.jpg)
Location example #3
(Demo)
![Page 30: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/30.jpg)
Questions now?
Have I managed to confuse anyone by now?
![Page 31: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/31.jpg)
Scalable geo hacks
Geo hacks -> scalable geo hacks.Keep in mind that everything we’ve talked about so far will scale well. I have “only” a few million location breadcrumbs — that’s a lot, but you might have many many more. The same index trees that got me this far are designed to keep going farther.
Now we’re going to talk about some hacks: some I haven’t actually used yet, the next makes your code uglier, and the last one adds cheating on top of that. But all of these “hacks” still have some good scaling properties.
![Page 32: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/32.jpg)
Hack #1
Geo: not just for geo
![Page 33: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/33.jpg)
Hack #1
n-dimensional
data
one-dimensional view
We talked about this: “projecting” an aspect of multi-dimensional data onto a 1-d sorted index. But why not project onto a...
![Page 34: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/34.jpg)
Hack #1
n-dimensional
datatwo-dimensional view!
2-d index!
![Page 35: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/35.jpg)
Hack #1
0
2750
5500
8250
11000
2008 2009 2010 2011 2012
Time and altitude, e.g. when did I fly in 2009–2010
![Page 36: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/36.jpg)
Hack #1
0
1
2
3
4
5
1 2 3 4
Camera and rating, e.g. clean up bad photos from recent cameras
![Page 37: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/37.jpg)
Hack #1
Consider:
?bbox=…&limit=500
?bbox=…&count=true
Unfortunately, this does not provide a *sorted* index. GeoCouch is scalable in the sense that the determining which objects are within the bounding box will stay fast, but the result set might be manageable only when counting or limiting.
![Page 38: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/38.jpg)
Hack #1
(No demo, sorry)
![Page 39: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/39.jpg)
Hack #2
All log(n) you can eat
![Page 40: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/40.jpg)
Hack #2
Map function (usual timestamp/location emit)
n = Math.log(values.length)while averages.length < n: averages.push(another)return averages
Reduce function
...
![Page 41: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/41.jpg)
Hack #3
Out-of-the ordinary location summary
![Page 42: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/42.jpg)
Hack #3
…?
Averaging locations puts dots on places I’ve never actually been.
![Page 43: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/43.jpg)
Hack #3
What might be more interesting is to pick good samples of actual locations I’ve been too. “Representative” locations so to speak. What’s a good “representative” location? The most common, ones with a lot of other points nearby, ones closest to that average we calculate?
![Page 44: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/44.jpg)
Hack #3
Finally realized that it was actually the “outliers” that provided the most interesting answer to “where have I all been”!
![Page 45: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/45.jpg)
Hack #3
Map function
n = Math.log(values.length)while outliers.length < n: outliers.push(another)return outliers
Reduce function
...
using log(n) trick from before
![Page 46: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/46.jpg)
Hack #3
Caution: RESULT IS “UNDEFINED”
This cheats.
Reduce function should be “commutative and associative for the array value input, to be able reduce on its own output and get the same answer”
This is not really associative: the result depends on the internal tree structure. The points picked are random/unstable.
![Page 47: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/47.jpg)
Hack #3
(but is interesting anyway)
In this case, I’d rather have randomly picked but practically useful output data, than stable “averages”.
![Page 48: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/48.jpg)
Hack #3
(Demo)
![Page 49: Scaling geodata w mapreduce natevw](https://reader030.fdocuments.in/reader030/viewer/2022032617/55a9a6b51a28abc2518b48c1/html5/thumbnails/49.jpg)
Thanks!(Any more questions?)
@natevwhttp://exts.ch