Craig Kerstiens - Scalable Uniques in Postgres @ Postgres Open
3 likes2,428 views
This document discusses using HyperLogLog in PostgreSQL to estimate cardinality or unique counts within a small memory footprint. It introduces HyperLogLog concepts like KMV, bit patterns, and stochastic averaging. It then demonstrates creating a PostgreSQL extension, inserting data into an HLL column, and using HLL functions to estimate unique counts across rows and tables. It also covers tuning HLL parameters and best practices like batching updates. The document presents HLL as a way to estimate large unique counts with only 1280 bytes and a few percent error rate.
![[Meetup] a successful migration from elastic search to clickhouse](https://cdn.slidesharecdn.com/ss_thumbnails/meetupasuccessfulmigrationfromelasticsearchtoclickhouse-191004114403-thumbnail.jpg?width=560&fit=bounds)
More Related Content
What's hot (20)
Viewers also liked (14)
Similar to Craig Kerstiens - Scalable Uniques in Postgres @ Postgres Open (20)
More from PostgresOpen (18)
Recently uploaded (20)
Craig Kerstiens - Scalable Uniques in Postgres @ Postgres Open
- 1. Scalable Uniques in Postgres - Craig Kerstiens Heroku Postgres
- 3. Truviso • Extended Postgres to do streaming • Various markets • Ad space • Wanted unique impressions • Sort of wanted unique impressions
- 5. Approx Top K
- 7. HyperLogLog
- 8. HyperLogLog • KMV - K minimum value
- 9. HyperLogLog • KMV - K minimum value • Bit observable patterns
- 10. HyperLogLog • KMV - K minimum value • Bit observable patterns • Stochastic averaging
- 11. HyperLogLog • KMV - K minimum value • Bit observable patterns • Stochastic averaging • Harmonic averaging
- 12. HyperLogLog • KMV - K minimum value • Bit observable patterns • Stochastic averaging • Harmonic averaging
- 13. HyperLogLog • KMV - K minimum value • Bit observable patterns • Stochastic averaging • Harmonic averaging • Implemented by Aggregate Knowledge
- 15. HyperLogLog Probabilistic uniques with small footprint
- 16. HyperLogLog Probabilistic uniques with small footprint Close enough distinct with small footprint
- 17. Use cases
- 18. Use cases • Semi distinct count • Think pg_stat_statements • Ad networks • Web traffic
- 19. Use cases • Semi distinct count • Think pg_stat_statements • Ad networks • Web traffic • With rollups/groupings
- 20. Digging in CREATE EXTENSION hll; CREATE TABLE helloworld ( id integer, set hll );
- 21. Digging in CREATE EXTENSION hll; CREATE TABLE helloworld ( id integer, set hll );
- 22. Inserting data UPDATE helloworld SET set = hll_add(set, hll_hash_integer(12345)) WHERE id = 1; UPDATE helloworld SET set = hll_add(set, hll_hash_text('hello world')) WHERE id = 1;
- 23. Real world CREATE TABLE daily_uniques ( date date UNIQUE, users hll );
- 24. Real world INSERT INTO daily_uniques(date, users) SELECT date, hll_add_agg(hll_hash_integer(user_id)) FROM users GROUP BY 1;
- 25. Real world SELECT EXTRACT(MONTH FROM date) AS month, hll_cardinality(hll_union_agg(users)) FROM daily_uniques WHERE date >= '2012-‐01-‐01' AND date < '2013-‐01-‐01' GROUP BY 1;
- 26. Real world SELECT EXTRACT(MONTH FROM date) AS month, hll_cardinality(hll_union_agg(users)) FROM daily_uniques WHERE date >= '2012-‐01-‐01' AND date < '2013-‐01-‐01' GROUP BY 1;
- 27. Good practices
- 28. Good practices
- 29. Good practices • It uses update
- 30. Good practices • It uses update • Do as a batch in most cases
- 31. Good practices • It uses update • Do as a batch in most cases • Tweak the config
- 33. Tuning Parameters • log2m - log base 2 of registers • Between 4 and 17 • Each 1 increase doubles storage
- 34. Tuning Parameters • log2m - log base 2 of registers • Between 4 and 17 • Each 1 increase doubles storage • regwidth - bits per register
- 35. Tuning Parameters • log2m - log base 2 of registers • Between 4 and 17 • Each 1 increase doubles storage • regwidth - bits per register • expthresh - threshold for explicit vs sparse
- 36. Tuning Parameters • log2m - log base 2 of registers • Between 4 and 17 • Each 1 increase doubles storage • regwidth - bits per register • expthresh - threshold for explicit vs sparse • spareson - on/off for sparse
- 37. Is it better?
- 38. 1280 bytes Estimate count of 10s of billions Few percent error
- 39. Resources • https://github.com/aggregateknowledge/ postgresql-hll • http://blog.aggregateknowledge.com/ 2013/02/04/open-source-release- postgresql-hll/ • http://tapoueh.org/blog/2013/02/25- postgresql-hyperloglog
- 40. Questions