Distributed COUNT(DISTINCT) with Hyperloglog on PostgreSQL | PGConf EU 2017 | Burak Yucesoy

Running SELECT COUNT(DISTINCT) on your database is all too common. In applications, it’s typical to have some analytics dashboard highlighting the number of unique items such as unique users or unique visits. While traditional SELECT COUNT(DISTINCT) queries works well in single machine setups, it is a difficult problem to solve in distributed systems. When you have this type of query, you cannot just push query to the workers and add up results, because most likely there will be overlapping records in different workers.

In this talk, we will focus on HyperLogLog(HLL) algorithm and its PostgreSQL extension postgresql-hll. HLL can provide approximate answers to COUNT(DISTINCT) queries in mathematically provable error bounds. It is not only fast and memory-efficient but also has very interesting properties which especially shine in distributed environment. During the talk, first, we’ll look at the internals of the HLL to understand why HLL algorithm is useful to solve distinct count problem in scalable way, then how it can be applied in a distributed fashion. Finally we will see some examples of HLL usage.

Transcript

1. Burak Yucesoy | Citus Data | PGConf EU Distributed COUNT(DISTINCT) with HyperLogLog on PostgreSQL
2. Burak Yucesoy | Citus Data | PGConf EU What is COUNT(DISTINCT)? • Number of unique elements (cardinality) in given data • Useful to find things like… ◦ Number of unique users visited your web page ◦ Number of unique products in your inventory
3. Burak Yucesoy | Citus Data | PGConf EU What is distributed COUNT(DISTINCT)? Worker Node 1 logins_001 Coordinator Worker Node 2 logins_002 Worker Node 3 logins_003
4. Burak Yucesoy | Citus Data | PGConf EU Why do we need distributed COUNT(DISTINCT)? • Your data is too big to fit in memory of single machine • Naive approach for COUNT(DISTINCT) needs too much memory
5. Burak Yucesoy | Citus Data | PGConf EU Why does distributed COUNT(DISTINCT) is difficult? Worker Node 1 logins_001 Coordinator SELECT COUNT(*) FROM logins; Worker Node 2 logins_002 Worker Node 3 logins_003 600 100 200 300 SELECT COUNT(*) FROM ...;
6. Burak Yucesoy | Citus Data | PGConf EU Why does distributed COUNT(DISTINCT) is difficult? Worker Node 1 logins_001 Coordinator SELECT COUNT(DISTINCT username) FROM logins; Worker Node 2 logins_002 Worker Node 3 logins_003 SELECT COUNT(DISTINCT user_id) FROM ...;
7. Burak Yucesoy | Citus Data | PGConf EU Why does distributed COUNT(DISTINCT) is difficult? Worker Node 1 logins_001 username | date ----------+----------- Alice | 2017-01-02 Bob | 2017-01-03 Charlie | 2017-01-05 Eve | 2017-01-07 Worker Node 3 logins_003 username | date ----------+----------- Frank | 2017-03-23 Eve | 2017-03-29 Charlie | 2017-03-02 Charlie | 2017-03-03 Worker Node 2 logins_002 username | date ----------+----------- Bob | 2017-02-11 Bob | 2017-02-13 Dave | 2017-02-17 Alice | 2017-02-19
8. Burak Yucesoy | Citus Data | PGConf EU Why does distributed COUNT(DISTINCT) is difficult? Worker Node 1 logins_001 username | date ----------+----------- Alice | 2017-01-02 Bob | 2017-01-03 Charlie | 2017-01-05 Eve | 2017-01-07 Worker Node 3 logins_003 username | date ----------+----------- Dave | 2017-03-23 Eve | 2017-03-29 Charlie | 2017-03-02 Charlie | 2017-03-03 Worker Node 2 logins_002 username | date ----------+----------- Bob | 2017-02-11 Bob | 2017-02-13 Dave | 2017-02-17 Alice | 2017-02-19
9. Burak Yucesoy | Citus Data | PGConf EU Some Possible Approaches • Pull all distinct data to one node and count there. (Doesn’t scale) • Repartition data on the fly. (Scales but it’s very slow) • Use HyperLogLog. (Scales and fast)
10. Burak Yucesoy | Citus Data | PGConf EU HyperLogLog(HLL) HLL is; • Approximation algorithm • Estimates cardinality of given data • Mathematically proven error bounds
11. Burak Yucesoy | Citus Data | PGConf EU Is it OK to approximate? It depends…
12. Burak Yucesoy | Citus Data | PGConf EU HLL • Very fast • Low memory footprint • Can work with streaming data • Can merge estimations of two separate datasets efficiently
13. Burak Yucesoy | Citus Data | PGConf EU How does HLL work? Steps; 1. Hash all elements a. Ensures uniform data distribution b. Can treat all data types same 2. Observing rare bit patterns 3. Stochastic averaging
14. Burak Yucesoy | Citus Data | PGConf EU How does HLL work? - Observing rare bit patterns hash Alice 645403841 binary 0010...001 Number of leading zeros: 2 Maximum number of leading zeros: 2
15. Burak Yucesoy | Citus Data | PGConf EU How does HLL work? - Observing rare bit patterns hash Bob 1492309842 binary 0101...010 Number of leading zeros: 1 Maximum number of leading zeros: 2
16. Burak Yucesoy | Citus Data | PGConf EU How does HLL work? - Observing rare bit patterns ... Maximum number of leading zeros: 7 Cardinality Estimation: 27
17. Burak Yucesoy | Citus Data | PGConf EU How does HLL work? Stochastic Averaging Measuring same thing repeatedly and taking average.
18. Burak Yucesoy | Citus Data | PGConf EU
19. Burak Yucesoy | Citus Data | PGConf EU
20. Burak Yucesoy | Citus Data | PGConf EU How does HLL work? Stochastic Averaging Data Partition 1 Partition 3 Partition 2 7 5 12 228.968... Estimation 27 25 212
21. Burak Yucesoy | Citus Data | PGConf EU How does HLL work? Stochastic Averaging 01000101...010 First m bits to decide partition number Remaining bits to count leading zeros
22. Burak Yucesoy | Citus Data | PGConf EU Error rate of HLL is damn good • Typical Error Rate: 1.04 / sqrt(number of partitions) • Memory need is number of partitions * log(log(max. value in hash space)) bit • Can estimate cardinalities well beyond 109 with 1% error rate while using a memory of only 6 kilobytes • Memory vs accuracy tradeoff
23. Burak Yucesoy | Citus Data | PGConf EU Why does HLL work? It turns out, combination of lots of bad estimation is a good estimation
24. Burak Yucesoy | Citus Data | PGConf EU Some interesting examples Alice Alice Alice … … … Alice Partition 1 Partition 3 Partition 2 0 2 0 1.103... Harmonic Mean 20 22 20 hash Alice 645403841 binary 00100110...001 ... ... ...
25. Burak Yucesoy | Citus Data | PGConf EU Some interesting examples Charlie Partition 1 Partition 8 Partition 2 29 0 0 1.142... Harmonic Mean 229 20 20 hash Charlie 0 binary 00000000...000 ... ... ...
26. Burak Yucesoy | Citus Data | PGConf EU postgresql-hll • https://github.com/aggregateknowledge/postgresql-hll • https://github.com/citusdata/postgresql-hll • Companies using postgresql-hll for their dashboard • Neustar • Cloudflare
27. Burak Yucesoy | Citus Data | PGConf EU postgresql-hll uses a data structure, also called hll to keep maximum number of leading zeros of each partition. • Use hll_hash_bigint to hash elements. ◦ There are some other functions for other common data types. • Use hll_add_agg to aggregate hashed elements into hll data structure. • Use hll_cardinality to materialize hll data structure to actual distinct count. postgresql-hll in single node
28. Burak Yucesoy | Citus Data | PGConf EU What Happens in Distributed Scenario?
29. Burak Yucesoy | Citus Data | PGConf EU How to merge COUNT(DISTINCT) with HLL Shard 1 Shard 1 Partition 1 Shard 1 Partition 3 Shard 1 Partition 2 7 5 12 HLL(7, 5, 12) Intermediate Result
30. Burak Yucesoy | Citus Data | PGConf EU How to merge COUNT(DISTINCT) with HLL Shard 2 Shard 2 Partition 1 Shard 2 Partition 3 Shard 2 Partition 2 11 7 8 HLL(11, 7, 8) Intermediate Result
31. Burak Yucesoy | Citus Data | PGConf EU How to merge COUNT(DISTINCT) with HLL 11 7 12 1053.255 211 27 212 HLL(11, 7, 8) HLL(7, 5, 12) HLL(11, 7, 12) hll_union_agg
32. Burak Yucesoy | Citus Data | PGConf EU How to merge COUNT(DISTINCT) with HLL Shard 1 + Shard 2 Shard 1 Partition 1(7) + Shard 2 Partition 1(11) 11 7 12 1053.255 Estimation Shard 1 Partition 2(5) + Shard 2 Partition 2(7) Shard 1 Partition 3(12) + Shard 2 Partition 4(8)
33. Burak Yucesoy | Citus Data | PGConf EU 1. Separate data into shards. postgresql-hll in distributed environment logins_001 logins_002 logins_003
34. Burak Yucesoy | Citus Data | PGConf EU 2. Put shards into separate nodes. postgresql-hll in distributed environment Worker Node 1 Coordinator Worker Node 2 Worker Node 3 logins_001 logins_002 logins_003
35. Burak Yucesoy | Citus Data | PGConf EU 3. For each shard, calculate hll (but do not materialize). postgresql-hll in distributed environment Shard 1 Shard 1 Partition 1 Shard 1 Partition 3 Shard 1 Partition 2 7 5 12 HLL(7, 5, 12) Intermediate Result
36. Burak Yucesoy | Citus Data | PGConf EU 4. Pull intermediate results to a single node. postgresql-hll in distributed environment Worker Node 1 logins_001 Coordinator Worker Node 2 logins_002 Worker Node 3 logins_003 HLL(6, 4, 11) HLL(10, 6, 7) HLL(7, 12, 5)
37. Burak Yucesoy | Citus Data | PGConf EU 5. Merge separate hll data structures and materialize them postgresql-hll in distributed environment 11 13 12 10532.571... 211 213 212 HLL(11, 7, 8) HLL(7, 5, 12) HLL(11, 13, 12) HLL(8, 13, 6)
38. Burak Yucesoy | Citus Data | PGConf EU Or use Citus :) postgresql-hll in distributed environment
39. Burak Yucesoy | Citus Data | PGConf EU Burak Yucesoy burak@citusdata.com @byucesoy Thank You citusdata.com | @citusdata