std.format is slow

[buremba]

We're heavily using string concentration via std.format and I realized that it dramatically hurts the performance. We had to switch using the old way '' + '' and had a huge performance boost. Here is an example:

["%s and %s" % ['test', 'test1'] for x in std.range(0, 1000)]

The script above is compiled in ~5000ms whereas the following one only takes around 50ms:

local test = 'test';
local test1 = 'test1';

[test + 'and' + test1 for x in std.range(0, 1000)]

I see that std.format is implemented in Jsonnet but I believe it might be best to implement it as a native function as the logic is complex and it affects the performance dramatically.

[sbarzowski]

Well, std.format will always be somewhat slower than just using +, because it needs to parse the strings format and then concatenate the parts. Of course it will be way way faster if implemented as a builtin.

BTW probably a large part of 50ms in the latter example is actually printing the output. The calculation itself is probably even faster than that.

[buremba]

@sbarzowski sure, that's expected. On the other hand, it would be nice to have a convenient way to concentrate the string other than using '' + '' if std.format is designed to be more powerful rather than being the default way to concentrate the strings. It's hard to read the '' + '' format and most programming languages provide a simple way such as '${variable}'.

Jsonnet doesn't need to cover all the cases but I believe that string concentration is a common use-case for a data templating language. WDYT?

P.S: For one of our projects, almost 80% of the time is spent on std.format when compiling the Jsonnet.

[CertainLach]

Rust implementation of jsonnet (https://github.com/CertainLach/jrsonnet/) implements std.format in native, and its performance is almost same as string concatenation (while still taking 1/2 of golang impl time)

# a.jsonnet
["%s and %s" % ['test', 'test1'] for x in std.range(0, 1000)]

# b.jsonnet
local test = 'test';
local test1 = 'test1';

[test + 'and' + test1 for x in std.range(0, 1000)]

Command	Mean [ms]	Min [ms]	Max [ms]	Relative
./target/release/jrsonnet a.jsonnet	2.9 ± 0.1	2.6	3.7	1.20 ± 0.09
./target/release/jrsonnet b.jsonnet	2.4 ± 0.1	2.2	4.4	1.00
jsonnet a.jsonnet	521.0 ± 2.2	518.3	524.6	217.80 ± 13.39
jsonnet b.jsonnet	4.2 ± 0.2	3.9	5.9	1.74 ± 0.13

[sbarzowski]

That's nice! This is the implementaiton, right? https://github.com/CertainLach/jrsonnet/blob/8db09e1bbe2ea8a499da56d2b972a5510a654f82/crates/jrsonnet-evaluator/src/builtin/format.rs

We might want to use it as an inspiration here.

[CertainLach]

Yep, it is