feat(export): add --analyze-config function-call profiler

[julienduchesne]

Summary

Add an opt-in Jsonnet function-call profiler to rtk export so it's easy to see which functions dominate an evaluation. Enable it with --analyze-config <file.json>, where the config selects how results are ranked and where they're written.

• output_file (required), plus optional sort_by (total (default), self, calls), total_time_threshold, and self_time_threshold (both in milliseconds, default 0)
• Report ranks functions and shows call count, inclusive total time, self time, and the defining file when known
• Zero overhead when the flag is absent: the hot path is gated by a single relaxed atomic load

Context

The single choke point for every Jsonnet call (FuncVal::evaluate) is wrapped with enter/exit timing. Each worker thread accumulates per-function counts and timings in thread-local storage and merges into a shared aggregate once per environment evaluation, so there's no per-call locking and the parallel export path stays fast.

Inclusive ("total") time is recursion-corrected: only the outermost frame of a given function on the stack contributes, so a recursive helper isn't counted once per stack level. Without this, a deeply recursive tokenizer showed a nonsensical 277s total (far above wall-clock); after the fix its caller correctly ranks above it and the numbers reflect the real call hierarchy.

Made with Cursor

[github-actions]

Benchmark Results

Benchmark	Test	vs tk	vs base
Diff	cluster_scoped	53.95x faster	equal
Diff	configmap_unchanged	32.16x faster	equal
Diff	deployment_nested_changes	30.54x faster	equal
Env List	Single Inline Directory (--json)	48.48x faster	equal
Env List	Single Inline File (--json)	47.62x faster	equal
Env List	All Environments (--json)	87.24x faster	equal
Eval	Single Static Environment	15.74x faster	equal
Eval	Inline Environment File	20.22x faster	equal
Eval	Eval Expression - Deeply Nested Path	15.63x faster	equal
Export (Full)	Single Static Environment	6.7x faster	equal
Export (Full)	Single Inline Environment	21.01x faster	equal
Export (Full)	All Environments	18.82x faster	equal
Export (Replace)	Single Static Environment	5.81x faster	equal
Export (Replace)	Single Inline Environment	20.77x faster	equal
Export (Replace)	All Environments	27.61x faster	equal
Helm Template	Full Export	1.73x faster	1.22x faster
Tool Importers Count	Single Lib Directory (Non-Recursive)	2.39x faster	equal
Tool Importers Count	Single Lib Directory (Recursive)	2.33x faster	equal
Tool Importers Count	Entire Lib Directory (Non-Recursive)	1.41x faster	equal
Tool Importers Count	Entire Lib Directory (Recursive)	1.12x faster	equal
Tool Importers	Global Lib File	3.14x faster	equal
Tool Importers	Single Env-Specific Lib File	1.59x faster	equal
Tool Importers	All Env-Specific Lib Files	5.29x faster	equal
Tool Importers	All Lib Files	5.08x faster	equal
Tool Importers	All Jsonnet Files	8.79x faster	equal
Tool Imports	Single Static Environment	3.28x faster	equal
Tool Imports	Single Inline Environment File	3.48x faster	equal
Validate Manifests	All Validations	-	equal
Validate Manifests	Manifest Tests Only	-	equal
Validate Manifests	Namespace Test Only	-	equal

Full results available in workflow artifacts.

Benchmark run on commit af3e9b4f194bfa08b317776316c41231be616680

[juliajohannesen]

Wouldn't this be better as an option that enables dynamic tracing spans? IE using the --analyze-config has the program start with a custom tracer that captures function call spans from jrsonnet?