Alex Osterloh (Google) - Big Data the Cloud Way

Big Data the Cloud Way

Transcript

1. Photo source: istockphoto.com Big Data - The Cloud Way
2. Photo source: istockphoto.com Big Data - The Cloud Way Alex Osterloh Google for Work aosterloh@google.com, linkedin.com/in/aosterloh
3. Google confidential │ Do not distribute Google confidential │ Do not distribute I dare you say “Big Data” one more time
4. Google confidential │ Do not distribute Google confidential │ Do not distribute 1 minute at Google scale Big Data @Google Agenda Big Data in the Cloud
5. Google confidential │ Do not distribute Google confidential │ Do not distribute 1 minute at Google scale But first ….
6. Photo source: istockphoto.com The future of Cloud Computing
7. The mainframe
8. The PC
9. The result
10. The mobile web
11. The mobile app
12. We often treat new technologies like those we already know before we realize their true potential. Disruption
13. The same thing is happening with cloud today. Disruption
14. Disruption The datacenter is not a collection of computers. The datacenter IS the computer.
15. Photo source: istockphoto.com Big Data @Google
16. “Organize the world’s information and make it universally accessible and useful.” Google’s Mission
17. Google confidential | Do not distribute Enterprise Google Trends Big Data is everywhere
18. Google confidential | Do not distribute Enterprise 1 minute at Google scale Big Data is everywhere
19. Google confidential | Do not distribute Enterprise 1 minute at Google scale Big Data is everywhere
20. 3 million searches 1000 new devices 100 hours 1 billion users 1 billion users 100 million gigabytes and also... 1 billion activated devices One minute at Google scale
21. Google confidential │ Do not distribute Google’s innovation around dealing with big data Spanner Dremel MapReduce Big Table Colossus 2012 2013 2002 2004 2006 2008 2010 GFS MillWheel Flume
22. Many Dremel use cases @Google • Analysis of crawled web documents • Crash reporting for Google products • OCR results from Google Books • Spam analysis • Debugging of map tiles on Google Maps • Results of tests run on Google’s distributed build system • Disk I/O statistics for hundreds of thousands of disks • Resource monitoring for jobs run in Google’s data centers
23. SELECT top(appId, 20) AS app, count(*) AS count FROM installlog.2012; ORDER BY count DESC Dremel Use @Google 1: Google Play How can a business analyst find Top 20 Apps in matter of seconds? Scan of ~1 Billion records Result in under 20 seconds
24. select date, rejection_reason, count(*) from line_item_table. last30days where line_item_id=56781234 Dremel Use @Google 2: DoubleClick Support What is root cause why ad was or was not delivered in the last 30 days? Scan of ~1.2B records Result in under 5 seconds
25. Dremel Architecture • Traffic minimization • Higher Compression ratio • Great for reads, not writes
26. Dremel Architecture Mixer 0 Mixer 1 Mixer 1 Leaf Leaf Leaf Leaf Distributed Storage • Columnar Storage • Long lived shared serving tree • Partial Reduction • Diskless data flow
27. Today on the Google toilet...
28. Photo source: istockphoto.com Big Data in the Cloud with BigQuery and Cloud Dataflow
29. Origins of BigQuery Google BigQuery http://bit.ly/dremelwp
30. Example: Using BigQuery to analyse 173M taxi trips SELECT INTEGER(AVG(tip_amount)*100)/100 avg_tip, REGEXP_EXTRACT(pickup_datetime, "2013- ([0-9]*)") month FROM [833682135931:nyctaxi.trip_fare] WHERE payment_type='CRD' GROUP BY 2 ORDER BY 1 desc Click Taxi to get to Live Demo What is the best month for getting tips as a NYC Taxi cab driver ?
31. Example: Using BigQuery to analyse 173M taxi trips Click Taxi to get to Live Demo What is the best month for getting tips as a NYC Taxi cab driver ? SELECT INTEGER(AVG(tip_amount)*100)/100 avg_tip, REGEXP_EXTRACT(pickup_datetime, "2013- ([0-9]*)") month FROM [833682135931:nyctaxi.trip_fare] WHERE payment_type='CRD' GROUP BY 2 ORDER BY 1 desc
32. Example: Using BigQuery to analyse 173M taxi trips Click Taxi to get to Live Demo What is the best month for getting tips as a NYC Taxi cab driver ? SELECT INTEGER(AVG(tip_amount)*100)/100 avg_tip, REGEXP_EXTRACT(pickup_datetime, "2013- ([0-9]*)") month FROM [833682135931:nyctaxi.trip_fare] WHERE payment_type='CRD' GROUP BY 2 ORDER BY 1 desc
33. Example: Github (290M entries) SELECT repository_language, count (repository_language) as pushes FROM [githubarchive:github.timeline] where type="PushEvent" and repository_pushed_at contains "2012" Group by repository_language order by pushes desc limit 10 Click Icon to get to Live Demo How many uploads to Github by Language 2012 vs. 2015?
34. SELECT repository_language, count (repository_language) as pushes FROM [githubarchive:github.timeline] where type="PushEvent" and and repository_pushed_at contains "2012" Group by repository_language order by pushes desc limit 10 Example: Github (290M entries) Click Icon to get to Live Demo How many uploads to Github by Language 2012 vs. 2015? 2012 2015 More examples at http://bigqueri.es/
35. BigQuery vs. MapReduce • MapReduce ◦ Flexible batch processing, multiple MR stages ◦ High overall throughput ◦ High latency • BigQuery ◦ Optimized for SQL queries ◦ Very low latency and almost real-time ◦ Great for trial-and-error and ad hoc queries
36. Store data with reliability, redundancy and consistency Go from data to meaning Quickly! At scale ... BigQuery
37. Google Confidential and Proprietary Example : Ezakus running 1000 hadoop Cluster
38. Google confidential | Do not distribute Google Cloud Platform Storage Cloud Storage Cloud SQL Cloud Datastore Compute Compute Engine App Engine App Services BigQuery Cloud Dataflow Heavy lifting: Google Heavy lifting: You
39. Google confidential | Do not distribute IaaS Example: Breaking the MinuteSort Record Compute Engine
40. Google confidential | Do not distribute Click to deploy ...
41. Google Confidential But analysing Big Data can be hard Image source: http://blog.mikiobraun.de/2014/10/parts-bug-no-car-big-data-infrastructure.html
42. “At Google we don’t really use MapReduce anymore. ”
43. Big Data publications at Google 2012 2013 MapReduce Spanner/F1 2003 2006 2007 2010 2011 GFS MillWheel Dremel BigQuery Big Table Cloud Datastore Paxos impl. 2004 FlumeJava Cloud Dataflow MapReduce Dremel BigQuery MillWheel FlumeJava Cloud Dataflow
44. A small example …
45. Prefix Suggestions #ar #argentina, #arugularocks, #argylesocks #arg #argentina, #argylesocks, #argonauts #arge #argentina, #argentum, #argentine Example: Auto completing #hashtags g.co/io14videos
46. {a->[apple, art, argentina], ar->[art, argentina, armenia],...} Count ExpandPrefixes Top(3) Write Read ExtractTags {a->(argentina, 5M), a->(armenia, 2M), …, ar-> (argentina, 5M), ar->(armenia, 2M), ...} {#argentina scores!, watching #armenia vs #argentina, my #art project, …} {argentina, armenia, argentina, art, ...} {argentina->5M, armenia->2M, art->90M, ...} Tweets Predictions
47. Count ExpandPrefixes Top(3) Write Read ExtractTags Tweets Predictions Pipeline p = Pipeline.create(); p.begin() p.run(); .apply(ParDo.of(new ExtractTags())) .apply(Top.largestPerKey(3)) .apply(Count.create()) .apply(ParDo.of(new ExpandPrefixes()) .apply(TextIO.Write.to(“gs://…”)); .apply(TextIO.Read.from(“gs://…”)) class ExpandPrefixes … { ... public void processElement(ProcessContext c) { String word = c.element().getKey(); for (int i = 1; i <= word.length(); i++) { String prefix = word.substring(0, i); c.output(KV.of(prefix, c.element())); } } }
48. Pipeline p = Pipeline.create(); p.begin() .apply(TextIO.Read.from(“gs://…”)) .apply(ParDo.of(new ExtractTags()) .apply(Count.create()) .apply(ParDo.of(new ExpandPrefixes()) .apply(Top.largestPerKey(3)) .apply(TextIO.Write.to(“gs://…”)); p.run(); Google Cloud Dataflow
49. Google Cloud Dataflow Optimize Schedule GCS GCS
50. Google Cloud Pub/Sub for reads and writes. time #ar* rank game begins armenia wins! #argyle #armeniarocks Age out old data #argentinagoal Let’s stream it …
51. Pipeline p = Pipeline.create(); p.begin() .apply(TextIO.Read.from(“gs://…”)) .apply(ParDo.of(new ExtractTags())) .apply(Count.create()) .apply(ParDo.of(new ExpandPrefixes()) .apply(Top.largestPerKey(3)) .apply(TextIO.Write.to(“gs://…”)); p.run();
52. Pipeline p = Pipeline.create(); p.begin() .apply(TextIO.Read.from(“gs://…”)) .apply(ParDo.of(new ExtractTags())) .apply(Count.create()) .apply(ParDo.of(new ExpandPrefixes()) .apply(Top.largestPerKey(3)) .apply(TextIO.Write.to(“gs://…”)); p.run();
53. Pipeline p = Pipeline.create(); p.begin() .apply(PubsubIO.Read.from(“input_topic”)) .apply(ParDo.of(new ExtractTags())) .apply(Count.create()) .apply(ParDo.of(new ExpandPrefixes()) .apply(Top.largestPerKey(3)) .apply(PubsubIO.Write.to(“output_topic”)); p.run();
54. Pipeline p = Pipeline.create(); p.begin() .apply(PubsubIO.Read.from(“input_topic”)) .apply(ParDo.of(new ExtractTags())) .apply(Count.create()) .apply(ParDo.of(new ExpandPrefixes()) .apply(Top.largestPerKey(3)) .apply(PubsubIO.Write.to(“output_topic”)); p.run();
55. Pipeline p = Pipeline.create(); p.begin() .apply(PubsubIO.Read.from(“input_topic”)) .apply(Bucket.by(SlidingWindows.of(60, MINUTES)) .apply(ParDo.of(new ExtractTags())) .apply(Count.create()) .apply(ParDo.of(new ExpandPrefixes()) .apply(Top.largestPerKey(3)) .apply(PubsubIO.Write.to(“output_topic”)); p.run();
56. Google Cloud Dataflow Streaming Optimize Streaming Schedule Pub/Sub Pub/Sub
57. Google Confidential A simple, yet powerful SDK for building highly parallelized & flexible data processing pipelines A fully managed service which optimizes, schedules and executes data processing pipelines Cloud Dataflow
58. Google Confidential Data arrives continuously, why wait to process it? Dataflow makes stream processing the new default. Batch available (with no code change) for when it’s appropriate (e.g. historical data reprocessing) Cloud Dataflow
59. Back to the NY Taxis …. now with Cloud Dataflow Problem: Mapping 340M pickup and drop off to NYC neighborhood locations by frequency.
60. Back to the NY Taxis …. now with Cloud Dataflow • Map 340M geo coordinates to NYC neighborhoods (342 Polygons) • Used BigQuery as input and output • Dataflow optimizes code by collapsing multiple logical passes into a single execution pass • Used Dataflow with 5 workers to run the pipeline • 25K records per second Interactive map at: http://nyctaximap.appspot.com/ Blog Post: http://bit.ly/nycdataflow Problem: Mapping 340M pickup and drop off to NYC neighborhood locations by frequency.
61. Want to know the details on Dataflow ? http://bit.ly/dataflowdeepdive
62. Photo source: istockphoto.com Thank you. Alex Osterloh - Google Cloud Platform Team aosterloh@google.com, linkedin.com/in/aosterloh