Big (enough) data

and strategies for distributed geoprocessing

@robinkraft

Data Lab @ World Resources Institute

Me - not an engineer

but still useful!

@robinkraft

WRI - not Twitter

But sometimes has lots of data

Big vs. small

What about big enough?

I know "big enough" when I see it

When standard tools just aren't enough ...

  • RAM
  • Disk space
  • Max out CPUs for days
  • Mystery crashes
  • Etc.

... but the big data toolkit is overkill

Awkward middle ground

GlobalForestWatch.org

Demo!

FORMA

"Near real-time" forest loss

One guiding principle

Simplicity > optimal

You don't start with simplicity, simplicity is the goal"

- @leafletjs #foss4g

1 hour

human cost = 400x AWS cost

FORMA = imagery + stats

Green to brown + fires = forest loss?

  • Pixel time series

  • Spatial joins

  • Spatial filters

  • Regressions

1-2 desktops

  • ArcGIS
  • Python
  • Stata/Numpy

How do you scale from 10k pixels to 100 billion?

Everything is a raster

row col val
0   0   10
0   1   20
1   0   30
1   1   40

Rasters are text

Everything is text

Hadoop loves text!

But Hadoop isn't simple

Hadoop without the hassle

+

+

Cascalog

MapReduce, without thinking MapReduce

(def query
  (let [data-src [[0 0 10]
                  [0 1 20]
                  [1 0 30]
                  [1 1 40]]]
    (<- [?row ?col ?new-val]
        (data-src ?row ?col ?val)
        (* 5 ?val :> ?new-val))
> (??- query)
0 0 50
0 1 100
1 0 150
1 1 200

Basic Cascalog

Simple join

pixel-src [row col val]:

0 0 10
0 1 20
1 0 30
1 1 40


country-src [row col country]:

0 0 "Japan"
0 1 "China"
1 0 "England"
1 1 "South Africa"
> (??- query)
0 0 "Japan10"
0 1 "China20"
1 0 "England30"
1 1 "South Africa40"
(def query
  (<- [?row ?col ?country-val]
    (pixel-src ?row ?col ?val)
    (country-src ?row ?col ?country)
    (str ?country ?val :> ?country-val)))

Aggregate count of hot fires, by country

fire-src [lat lon date kelvin]:

30 -120 "2014-09-10" 335
39 116 "2014-09-10" 335
6 106 "2014-08-25" 300
6 106 "2014-08-25" 350
6 105 "2014-09-09" 339

country-src [row col country]:

300 400 "United States"
250 2000 "China"
200 2000 "Indonesia"
> (??- fires-query)
"China" 1
"Indonesia" 2
"United States" 1
(:use 'cascalog.api)
(:require '[cascalog.ops :as c])
(:use '[demo :only (latlon->row-col)])

(def fires-query
  (<- [?country ?count]
    (fire-src ?lat ?lon ?date ?kelvin)
    (country-src ?row ?col ?country)
    (latlon->row-col ?lat ?lon :> ?row ?col)
    (< 330 ?kelvin)
    (c/count ?count)))

Pixel time series

pixel-src [row col date val]:

0 0 "2010-01-01" 10
0 0 "2010-01-02" 20
0 0 "2010-01-03" 30
0 0 "2010-01-04" 40
0 1 "2010-01-01" 50
0 1 "2010-01-02" 60
0 1 "2010-01-03" 70
0 1 "2010-01-04" 80
(use 'cascalog.api)
(use '[demo :only (build-series)])

(def timeseries-query
  (<- [?row ?col ?start ?end ?series]
    (pixel-src ?row ?col ?date ?val)
    (build-series ?date ?val :> ?start ?end ?series)))
> (??- timeseries-query)
0 0 "2010-01-01" "2010-01-04" [10 20 30 40]
0 1 "2010-01-01" "2010-01-04" [50 60 70 80]

"Vector tiles" - SQL

INSERT INTO gfw2_forma (x,y,date_array,z)
     (SELECT x, y, array_agg(undate) AS date_array, 16 AS z
FROM (SELECT floor(x/2) AS x, floor(y/2) AS y, unnest(date_array) AS undate
      FROM gfw2_forma WHERE z = 16) foo
GROUP BY x,y)"
{
    "rows": [
        {
            "x": 51480,
            "y": 32760,
            "sd": [
                86,
                98
            ],
            "se": [
                1,
                1
            ]
        },
etc.
}

It works, but ...

  • Hard to test
  • Statement timeouts
  • Hard to automate
  • Slow for large table
(defn zoom-out
  [n]
  (math/floor (/ n 2)))

(defn gen-tiles
  [x y z min-z]
  (if (= z min-z)
     [[x y z]]
     (conj (gen-tiles (zoom-out x) (zoom-out y) (dec z) min-z) [x y z])))

(defn prep-xyz
  "Stylized version"
  [src]
  (<- [?x2 ?y2 ?z2 ?period ?count]
      (src ?row ?col ?start-period ?series)
      (gen-period ?start-period ?series :> ?period _)
      (rowcol->latlon ?row ?col :> ?lat ?lon)
      (latlon->tile ?lat ?lon zoom :> ?x ?y ?z)
      (c/count ?count)))

(defn gen-all-zooms
  [src min-z]
  (<- [?x2 ?y2 ?z2 ?period ?count]
      (src ?x ?y ?z ?period)
      (gen-tiles ?x ?y ?z min-z :> ?x2 ?y2 ?z2)
      (c/count ?count)))

> (let [src (hfs-seqfile "s3n://path/to/data"
        min-z 7]
    (??- (gen-all-zooms (prep-xyz src))))

Gen tiles for 10 pixels or 100 billion

Calculate tile for one pixel, all zooms

Transpose time series, convert rowcol -> latlon -> xyz

Run it!

Same result, but

  • "infinitely scalable"

  • testable

  • totally reliable

  • fast enough

Lessons for

"big enough" geo data

Find - and use! - the right tools

  • Hadoop
  • StarCluster
  • Spark
  • Python multiprocessing
  • AWS Simple Queue Service
  • Google App Engine Task Queues
  • GeoTrellis

Simple > optimal

1 hour of human = 400 AWS hours

Get creative

... about data formats
... about what geo means

Hadoop can be your friend, or your enemy ...

Questions?

 

@robinkraft

 

GlobalForestWatch.org

 

http://slides.com/wri/big-enough-foss4g

Big (enough) data and strategies for distributed geoprocessing

By wri

Big (enough) data and strategies for distributed geoprocessing

  • 953
Loading comments...

More from wri