1:05 in PostgreSQL 16 but it was harder than I thought to saturate the CPUs and not be disk-bound. Also, I ran on GCP not Hetzner, so maybe different hardware.
SQL in theory makes this trivial, handles many of the big optimizations and looking at the repo, cuts 1000+ LOC down to a handful. Modern SQL engines handle everything for you, which is the whole damned point of SQL. Any decent engine will handle parallelism, caching, I/O, etc. Some exotic engines can leverage GPU but for N=1 billion, I doubt GPU will be faster. Here's the basic query:
SELECT city, MIN(temp), AVG(temp), MAX(temp) FROM temps GROUP BY 1 ORDER BY 1;
The first optimization is to observe that the cardinality of cities is small and can be normalized into integers (INTEGER aka INT4), and that the temps can as well (1 decimal of precision). Using SMALLINT (aka INT2) is probably not faster on modern CPUs but should use less RAM, which is better for both caching on smaller systems and cache hitrate on all systems. NUMERIC generally isn't faster or tighter on most engines.
To see the query plan, use EXPLAIN:
postgres=# explain SELECT city, MIN(temp), AVG(temp), MAX(temp) FROM temps_int2 GROUP BY 1 ORDER BY 1 limit 5;
QUERY PLAN
---------------------------------------------------------------------------------------------------------------
Limit (cost=13828444.05..13828445.37 rows=5 width=38)
-> Finalize GroupAggregate (cost=13828444.05..13828497.22 rows=200 width=38)
Group Key: city
-> Gather Merge (cost=13828444.05..13828490.72 rows=400 width=38)
Workers Planned: 2
-> Sort (cost=13827444.02..13827444.52 rows=200 width=38)
Sort Key: city
-> Partial HashAggregate (cost=13827434.38..13827436.38 rows=200 width=38)
Group Key: city
-> Parallel Seq Scan on temps_int2 (cost=0.00..9126106.69 rows=470132769 width=4)
JIT:
Functions: 8
Options: Inlining true, Optimization true, Expressions true, Deforming trueSigh, pg16 is still pretty conservative about parallelism, so let's crank it up.
SET max_parallel_workers=16; set max_parallel_workers_per_gather=16;
SET min_parallel_table_scan_size=0; set min_parallel_index_scan_size=0;
SET parallel_setup_cost = 0; -- Reduce the cost threshold for parallel execution
SET parallel_tuple_cost = 0.001; -- Lower the cost per tuple for parallel execution
postgres=# explain SELECT city, MIN(temp), AVG(temp), MAX(temp) FROM temps_int2 GROUP BY 1 ORDER BY 1 limit 5;
...
Workers Planned: 14
...
top - 10:09:48 up 22 min, 3 users, load average: 5.36, 1.87, 0.95
Tasks: 169 total, 1 running, 168 sleeping, 0 stopped, 0 zombie
%Cpu(s): 12.6 us, 4.8 sy, 0.0 ni, 7.8 id, 74.2 wa, 0.0 hi, 0.7 si, 0.0 st
MiB Mem : 32084.9 total, 258.7 free, 576.7 used, 31249.5 buff/cache
MiB Swap: 0.0 total, 0.0 free, 0.0 used. 30902.4 avail Mem
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
1062 postgres 20 0 357200 227952 197700 D 17.9 0.7 9:35.88 postgres: 16/main: postgres postgres [local] SELECT
1516 postgres 20 0 355384 91232 62384 D 17.6 0.3 0:08.79 postgres: 16/main: parallel worker for PID 1062
1522 postgres 20 0 355384 93336 64544 D 17.6 0.3 0:08.53 postgres: 16/main: parallel worker for PID 1062
1518 postgres 20 0 355384 90148 61300 D 17.3 0.3 0:08.53 postgres: 16/main: parallel worker for PID 1062
1521 postgres 20 0 355384 92624 63776 D 17.3 0.3 0:08.54 postgres: 16/main: parallel worker for PID 1062
1519 postgres 20 0 355384 90440 61592 D 16.6 0.3 0:08.58 postgres: 16/main: parallel worker for PID 1062
1520 postgres 20 0 355384 92732 63884 D 16.6 0.3 0:08.49 postgres: 16/main: parallel worker for PID 1062
1517 postgres 20 0 355384 91544 62696 D 16.3 0.3 0:08.55 postgres: 16/main: parallel worker for PID 1062
OK enough pre-optimization, here's the baseline:
postgres=# SELECT city, MIN(temp), AVG(temp), MAX(temp) FROM temps_int2 GROUP BY 1 ORDER BY 1 limit 5;
city | min | avg | max
------+-----+----------------------+------
0 | 0 | 276.0853550961625011 | 1099
1 | 0 | 275.3679265859715333 | 1098
2 | 0 | 274.6485567539599619 | 1098
3 | 0 | 274.9825584419741823 | 1099
4 | 0 | 275.0633718875598229 | 1097
(5 rows)
Time: 140642.641 ms (02:20.643)
Instead, to amortize the tuple overhead, we can store the data as arrays:
CREATE TABLE temps_by_city AS SELECT city, array_agg(temp) from temps_int2 group by city;
$ ./table_sizes.sh
...total... | 36 GB
temps_int2 | 34 GB
temps_by_city | 1980 MB
-- https://stackoverflow.com/a/18964261/430938 adding IMMUTABLE PARALLEL SAFE
CREATE OR REPLACE FUNCTION array_min(_data ANYARRAY) RETURNS NUMERIC AS $$
SELECT min(a) FROM UNNEST(_data) AS a
$$ LANGUAGE SQL IMMUTABLE PARALLEL SAFE;
SET max_parallel_workers=16; set max_parallel_workers_per_gather=16;
SET min_parallel_table_scan_size=0; set min_parallel_index_scan_size=0;
SET parallel_setup_cost = 0; -- Reduce the cost threshold for parallel execution
SET parallel_tuple_cost = 0.001; -- Lower the cost per tuple for parallel execution
postgres=# create table tmp1 as select city, min(array_min(array_agg)), avg(array_avg(array_agg)), max(array_max(array_agg)) from temps_by_city group by 1 order by 1 ;
SELECT 1001
Time: 132616.944 ms (02:12.617)
postgres=# explain select city, min(array_min(array_agg)), avg(array_avg(array_agg)), max(array_max(array_agg)) from temps_by_city group by 1 order by 1 ;
QUERY PLAN
-------------------------------------------------------------------------------------------------
Sort (cost=338.31..338.81 rows=200 width=98)
Sort Key: city
-> Finalize HashAggregate (cost=330.14..330.66 rows=200 width=98)
Group Key: city
-> Gather (cost=321.52..322.64 rows=600 width=98)
Workers Planned: 3
-> Partial HashAggregate (cost=321.52..322.04 rows=200 width=98)
Group Key: city
-> Parallel Seq Scan on temps_by_city (cost=0.00..0.04 rows=423 width=34)Ah, only using 3 cores of the 8...
---
CREATE TABLE temps_by_city3 AS SELECT city, temp % 1000, array_agg(temp) from temps_int2 group by 1,2;
postgres=# explain select city, min(array_min(array_agg)), avg(array_avg(array_agg)), max(array_max(array_agg)) from temps_by_city3 group by 1 order by 1 ;
QUERY PLAN
---------------------------------------------------------------------------------------------------------
Finalize GroupAggregate (cost=854300.99..854378.73 rows=200 width=98)
Group Key: city
-> Gather Merge (cost=854300.99..854348.73 rows=2200 width=98)
Workers Planned: 11
-> Sort (cost=854300.77..854301.27 rows=200 width=98)
Sort Key: city
-> Partial HashAggregate (cost=854290.63..854293.13 rows=200 width=98)
Group Key: city
-> Parallel Seq Scan on temps_by_city3 (cost=0.00..98836.19 rows=994019 width=34)
JIT:
Functions: 7
Options: Inlining true, Optimization true, Expressions true, Deforming trueThis 100% saturates the CPU and runs in ~65 secs (about 2x faster).
---
Creating the data
Here's a very fast lousy first pass for PostgreSQL that works on most versions - for pg16, there's random_normal(). I'm not on Hetzner so I used GCP (c2d-standard-8 with 16vCPU, 64GB, and 150GB disk, ubuntu 22.04 and
CREATE TABLE temps_int2 (city int2, temp int2);
-- random()*random() is a cheap distribution. 1100 = 110 * 10 to provide one decimal of precision.
INSERT INTO temps_int2 (city, temp) SELECT (1000*random())::int2 as city, (random()*random()*1100)::int2 temp from generate_series(1,1e9)i;