Use 64bit floats for geoDistance calculation

[dadebue]

Describe the unexpected behaviour

• The geoDistance function (and the other geographical distance functions) only uses Float32 values for the calculation
• This doesn't provide a high enough precision for high precision positioning systems (like RTK GPS) which are able to determine the position up to ±1cm precision
• When using the geoDistance function with close enough positions the result is always 0, which is incorrect

How to reproduce

• most recent Clickhouse version
• SELECT geoDistance(8.623000, 49.355000, 8.6230001, 49.3550001); returns 0
• The exact distance is 0.01327m (link to online calculation)
  • Latitude 1: 49.3550001
  • Latitude 2: 49.3550000
  • Longitude 1: 8.6230001
  • Longitude 1: 8.6230000
• The exact distance can be calculated with Clickhouse too when using the full haversine formula manually (returns 0.01327037431348616):

SELECT (2*atan2(sqrt((sin((49.3550001-49.3550000)*pi()/360)*sin((49.3550001-49.3550000)*pi()/360))+(cos(49.3550001*pi()/180)*cos(49.3550000*pi()/180)*sin((8.6230001-8.6230000)*pi()/360)*sin((8.6230001-8.6230000)*pi()/360))),sqrt(1-(sin((49.3550001-49.3550000)*pi()/360)*sin((49.3550001-49.3550000)*pi()/360))+(cos(49.3550001*pi()/180)*cos(49.3550000*pi()/180)*sin((8.6230001-8.6230000)*pi()/360)*sin((8.6230001-8.6230000)*pi()/360)))))*6371000 distance;

Expected behavior

• The geoDistance function should use Float64 values for calculation to support high precision position values
• SELECT geoDistance(8.623000, 49.355000, 8.6230001, 49.3550001); should return 0.01327

Additional context

• Changes were introduced in Geo distance functions improve performance #37524
• I'm no c++ expert but I suppose those lines in src/Functions/greatCircleDistance.cpp need to be changed

[alexey-milovidov]

Let's switch to Float64 and check if there will be any noticeable difference in performance.
I thought there would be no difference.

[dadebue]

BTW this chart clearly shows the difference between the geoDistance and manual haversine formula and the lack of precision...

`

[geetptl]

I'd like to work on this.

[alexey-milovidov]

@geetptl, thank you! This will be very helpful.

[Uzair-90]

I'm interested in working on this issue. It aligns well with my skills and interests. I've reviewed the problem, and I believe I can contribute a solution.

[alexey-milovidov]

@Uzair-90, @geetptl, I don't know who will propose a working solution first, but I'd appreciate your work!

[geetptl]

Unsure how to link this PR here, but here's the link to it:

#58847

[Uzair-90]

@alexey-milovidov Thank you for the opportunity I will give it a try and highly appreciate it.

[dadebue]

@alexey-milovidov What do you think about the PR from geetptl #58847?

[alexey-milovidov]

@dadebue, it is unfinished - there is performance degradation.

[alexey-milovidov]

#61848

[alexey-milovidov]

@dadebue, @Uzair-90, @geetptl, I've reimplemented it. As a result, the performance increased instead of being degraded as expected before. Also, I've managed to make it 100% compatible and even allow the user to control the behavior.

https://s3.amazonaws.com/clickhouse-test-reports/61848/c2209c997ca08466a595b3ff9fec6db78e244123/performance_comparison_[2_4]/report.html

[geetptl]

@alexey-milovidov It was vacation when I was working on this issue, but student life caught up to me, along with my inexperience in C++. But thank you for your guidance in the now closed PR!

[alexey-milovidov]

Thank you, and it is actually your contribution :) I will add you as a co-author.