My SEA Amsterdam talk

Transcript

1. GoDataDriven PROUDLY PART OF THE XEBIA GROUP Real time data driven applications Giovanni Lanzani Data Whisperer Using Python + pandas as back end
2. Who am I? 2008-2012: PhD Theoretical Physics 2012-2013: KPMG 2013-Now: GoDataDriven
3. Feedback @gglanzani
4. Real-time, data driven app? •No store and retrieve; •Store, {transform, enrich, analyse} and retrieve; •Real-time: retrieve is not a batch process; •App: something your mother could use: SELECT attendees FROM pydataberlin2014 WHERE password = '1234';
5. Get insight about event impact
6. Get insight about event impact
7. Get insight about event impact
8. Get insight about event impact
9. Is it Big Data? •Raw logs are in the order of 40TB; •We use Hadoop for storing, enriching and pre- processing; •(10 nodes, 24TB per nodes)
10. Challenges 1. Privacy; 2. Huge pile of data; 3. Real-time retrieval; 4. Some real-time analysis.
11. 1. Privacy
12. 3. Real-time retrieval •Harder than it looks; •Large data; •Retrieval is by giving date, center location + radius.
13. 4. (Some) real-time analysis
14. Architecture AngularJS app.py helper.py REST Front-end Back-end JSON
15. JS - 1
16. JS - 2
17. Flask from flask import Flask app = Flask(__name__) @app.route("/hello") def hello(): return "Hello World!" if __name__ == "__main__": app.run()
18. app.py example @app.route('/api/<postcode>/<date>/<radius>', methods=['GET']) @app.route('/api/<postcode>/<date>', methods=['GET']) def datapoints(postcode, date, radius=1.0): ... stats, timeline, points = helper.get_json(postcode, date, radius) return … # returns a JSON object for AngularJS
19. data example date hour id_activity postcode hits delta sbi 2013-01-01 12 1234 1234AB 35 22 1 2013-01-08 12 1234 1234AB 45 35 1 2013-01-01 11 2345 5555ZB 2 1 2 2013-01-08 11 2345 5555ZB 55 2 2
20. Levenshtein distance def levenshtein(s, t): if len(s) == 0: return len(t) if len(t) == 0: return len(s) cost = 0 if s[-1] == t[-1] else 1 return min(levenshtein(s[:-1], t) + 1, levenshtein(s, t[:-1]) + 1, levenshtein(s[:-1], t[:-1]) + cost) levenshtein("search", “engine”) # 6 levenshtein("Amsterdam", “meetup") # 7 levenshtein("Lplein", “Leidseplein") # 5
21. Aside •Memoize every recursive algorithm in Python! from cytoolz import functoolz @functoolz.memoize def memo_levenshtein(s, t): if len(s) == 0: return len(t) if len(t) == 0: return len(s) cost = 0 if s[-1] == t[-1] else 1 return min(levenshtein(s[:-1], t) + 1, levenshtein(s, t[:-1]) + 1, levenshtein(s[:-1], t[:-1]) + cost)
22. Aside •Memoize the algorithm if you’re using Python! a = “Marco Polo str.” b = “Marco Polo straat” %timeit memo_levenshtein(a, b) # 213ns %timeit levenshtein(a, b) # 1.84 µs
23. helper.py example def get_json(postcode, date, radius): ... lat, lon = get_lat_lon(postcode) postcodes = get_postcodes(postcode, radius) data = get_data(postcodes, dates) stats = get_statistics(data, sbi) timeline = get_timeline(data, sbi) return stats, timeline, data.to_json(orient='records')
24. helper.py example def get_statistics(data, sbi): sbi_df = data[data.sbi == sbi] # select * from data where sbi = sbi hits = sbi_df.hits.sum() # select sum(hits) from … delta_hits = sbi_df.delta.sum() # select sum(delta) from … if delta_hits: percentage = (hits - delta_hits) / delta_hits else: percentage = 0 return {"sbi": sbi, "total": hits, "percentage": percentage}
25. helper.py example def get_timeline(data, sbi): df_sbi = data.groupby([“date”, “hour", “sbi"]).aggregate(sum) # select sum(hits), sum(delta) from data group by date, hour, sbi return df_sbi
26. helper.py example def get_json(postcode, date, radius): ... lat, lon = get_lat_lon(postcode) postcodes = get_postcodes(postcode, radius) dates = date.split(';') data = get_data(postcodes, dates) stats = get_statistics(data) timeline = get_timeline(data, dates) return stats, timeline, data.to_json(orient='records')
27. Who has my data? •First iteration was a (pre)-POC, less data (3GB vs 500GB); •Time constraints; •Oeps: import pandas as pd ... source_data = pd.read_csv("data.csv", …) ... def get_data(postcodes, dates): result = filter_data(source_data, postcodes, dates) return result
28. Advantage of “everything is a df” Pro: •Fast!! •Use what you know •NO DBA’s! •We all love CSV’s! Contra: •Doesn’t scale; •Huge startup time; •NO DBA’s! •We all hate CSV’s!
29. If you want to go down this path •Set the dataframe index wisely; •Align the data to the index: •Beware of modifications of the original dataframe! source_data.sort_index(inplace=True)
30. If you want to go down this path The reason pandas is faster is because I came up with a better algorithm
31. New architecture data = get_data(db, postcodes, dates) data = get_data(postcodes, dates) database.py Data psycopg2 AngularJS app.py helper.py REST Front-end Back-end JSON
32. Handling geo-data def get_json(postcode, date, radius): """ ... """ lat, lon = get_lat_lon(postcode) postcodes = get_postcodes(postcode, radius) dates = date.split(';') data = get_data(postcodes, dates) stats = get_statistics(data) timeline = get_timeline(data, dates) return stats, timeline, data.to_json(orient='records')
33. Issues?! •With a radius of 10km, in Amsterdam, you get 10k postcodes. You need to do this in your SQL: •Index on date and postcode, but single queries running more than 20 minutes. SELECT * FROM datapoints WHERE date IN date_array AND postcode IN postcode_array;
34. Postgres + Postgis (2.x) PostGIS is a spatial database extender for PostgreSQL. Supports geographic objects allowing location queries SELECT * FROM datapoints WHERE ST_DWithin(lon, lat, 1500) AND dates IN ('2013-02-30', '2013-02-31'); -- every point within 1.5km -- from (lat, lon) on imaginary dates
35. Other db’s?
36. Steps to solve it 1. Align data on disk by date; 2. Use the temporary table trick: 3. Lose precision: 1234AB→1234 CREATE TEMPORARY TABLE tmp (postcodes STRING NOT NULL PRIMARY KEY); INSERT INTO tmp (postcodes) VALUES postcode_array; SELECT * FROM tmp JOIN datapoints d ON d.postcode = tmp.postcodes WHERE d.dt IN dates_array;
37. Take home messages 1. Geospatial problems are “hard” and can kill your queries; 2. Not everybody has infinite resources: be smart and KISS! 3. SQL or NoSQL? (Size, schema)
38. GoDataDriven We’re hiring / Questions? / Thank you! @gglanzani giovannilanzani@godatadriven.com Giovanni Lanzani Data Whisperer