perf: std.manifestJsonEx is 1.74× slower than jrsonnet

[He-Pin]

Tracking issue for a specific perf gap found while comparing sjsonnet (native, master) against jrsonnet (master). Parent comparison: #666.

Observation

std.manifestJsonEx is 1.74× slower than jrsonnet.

Scenario: bench/resources/go_suite/manifestJsonEx.jsonnet.

	mean	min
sjsonnet (native)	4.7 ± 0.9 ms	3.3 ms
jrsonnet	2.7 ± 0.9 ms	1.5 ms

Repro:

hyperfine --warmup 2 --runs 20 -N \
  "sjsonnet bench/resources/go_suite/manifestJsonEx.jsonnet" \
  "jrsonnet bench/resources/go_suite/manifestJsonEx.jsonnet"

Code

sjsonnet/src/sjsonnet/stdlib/ManifestModule.scala:76-105 — ManifestJsonEx. Calls Materializer.apply0(v, MaterializeJsonRenderer(...))(ev).toString.

Hypothesis

Two-stage rendering:

1. Materializer walks the Val tree and feeds a ujson visitor (which builds a ujson.Value or streams to a writer depending on the renderer).
2. Output is collected into a StringWriter, then .toString copies the buffer into a new String, then Val.Str wraps it.

Each visitor call is a virtual dispatch. For the nested test object this is ~40–50 visits + indent logic.

jrsonnet appears to render directly from its Val to a String/IStr in a single pass, without an intermediate visitor AST.

Directions

• Specialize ManifestJsonEx to walk the Val tree directly and emit to a single CharBuilder / StringBuilder — skipping the Materializer + ujson indirection entirely. The main tradeoff is code duplication with the renderer — could be mitigated by sharing escape/number utilities.
• Since kube-prometheus (real-world) ends up materializing large JSON, this gap likely contributes to the ~2.8× real-world gap identified in performance optimization #666. High-impact.

Part of the jrsonnet-parity effort tracked in #666.

[He-Pin]

Closing — gap is inside hyperfine's sub-5 ms calibration band on current master

Refreshed on sjsonnet ca61a7a3 vs jrsonnet db1fd8f1:

	sjsonnet-native (ms)	jrsonnet (ms)	Ratio sj/jr
bench/resources/go_suite/manifestJsonEx.jsonnet	3.8 ± 0.3	2.4 ± 1.3	1.58×

Both sides are under 5 ms — hyperfine raises Command took less than 5 ms to complete. Note that the results might be inaccurate because hyperfine can not calibrate the shell startup time much more precise than this limit. The jrsonnet-side stddev (±1.3 ms) is in the same range as the absolute delta (~1.4 ms). PR #839 (reuse manifest JSON visitors) already landed on the sjsonnet side.

Closing as noise-bound. Will reopen if a larger-scale manifest benchmark (or real workload) surfaces a signal outside the calibration band.