fix: align error-message casing and punctuation with go-jsonnet by He-Pin · Pull Request #1007 · databricks/sjsonnet

He-Pin · 2026-06-20T05:05:07Z

Summary

Several sjsonnet error messages diverged from go-jsonnet's exact wording
in casing or punctuation. Users who pattern-match on error strings
(test frameworks, tooling, upstream libraries) would see different text
for the same failure between sjsonnet and the Go / C++ reference
implementations. This PR batches the text fixes into one PR.

Modification

std.pow / std.asin / std.acos NaN: capitalize "not a number"
→ "Not a number" to match go-jsonnet's makeDoubleCheck output and
the convention already used by std.sqrt, std.log, std.log2,
std.log10 in sjsonnet (MathModule.scala).
Val.Num.asDouble NaN: capitalize "not a number" → "Not a number"
so the asDouble accessor matches the same convention (Val.scala).
Val.Num constructor Infinity: capitalize "overflow" → "Overflow"
so Val.Num's infinity check matches go-jsonnet's makeDoubleCheck
convention (Val.scala).
Arithmetic-operator NaN / Infinity in Evaluator.scala: capitalize
all 12 "not a number" / "overflow" errors to "Not a number" /
"Overflow" so operators and builtins emit consistent casing.
std.avg([]) empty: append the missing trailing period to
"Cannot calculate average of an empty array" → "Cannot calculate average of an empty array."
to match the exact wording used by go-jsonnet and C++ jsonnet
(ArrayModule.scala).
Regenerated golden files to reflect the new wording:
error.math_pow_nan / error.math_asin_nan / error.math_acos_nan
(new_test_suite); pow4, pow7, builtin_exp3, builtin_exp5,
builtin_log5, div4, inf_min_number, inf_mul_number,
inf_sum_number (go_test_suite); error.arithmetic_overflow_*
(3 files, new_test_suite); error.overflow2 (test_suite).
Updated 3 test assertions: StdMathTests.scala (log(0) / log2(0)
overflow assertions) and TailCallOptimizationTests.scala
(tailstrictFactorialOverflow) — all changed from "overflow" to
"Overflow".

Result

sjsonnet error messages for these failure modes now byte-match
go-jsonnet and C++ jsonnet (or deviate only in documented, minor,
error-text-only ways). All file tests (new_test_suite +
go_test_suite + test_suite), StdMathTests, and
TailCallOptimizationTests pass on Scala 2.12 / 2.13 / 3.3.

Cross-implementation comparison (final state)

Failure mode	sjsonnet (before)	sjsonnet (after)	go-jsonnet 0.22.0	jrsonnet 0.5.0-pre99	C++ jsonnet 0.22.0
`std.pow(-1, 0.5)` NaN	`"not a number"`	`"Not a number"`	`"Not a number"`	`"non-finite"`	`"not a number"`
`std.asin(2)` NaN	`"not a number"`	`"Not a number"`	`"Not a number"`	`"non-finite"`	`"not a number"`
`std.acos(2)` NaN	`"not a number"`	`"Not a number"`	`"Not a number"`	`"non-finite"`	`"not a number"`
`std.sqrt(-1)` NaN	`"Not a number"` (already correct)	`"Not a number"`	`"Not a number"`	`"out of bounds"`	`"not a number"`
`0.0 / 0.0` NaN (operator)	`"not a number"`	`"Not a number"`	`"Not a number"`	`"non-finite"`	`"not a number"`
`std.exp(1000)` Infinity	`"overflow"`	`"Overflow"`	`"Overflow"`	`"non-finite"`	`"overflow"`
`1e308 * 2` Infinity (op)	`"overflow"`	`"Overflow"`	`"overflow"` *	`"non-finite"`	`"overflow"` *
`std.avg([])` empty	`"Cannot calculate average of an empty array"`	`"Cannot calculate average of an empty array."`	`"Cannot calculate average of an empty array."`	`"expected non-empty array"`	`"Cannot calculate average of an empty array."`

* go-jsonnet / C++ jsonnet keep lowercase "overflow" for operator-level
overflows. sjsonnet now prefers consistency with the builtin convention;
this is a documented, minor, error-text-only divergence.

Intentionally deferred

std.assertEqual message format: The message lives in stdlib.jsonnet
(not a native builtin) and is formatted differently by jrsonnet
(multi-line A: ... / B: ... form). Changing it would require
stdlib-level edits and affect many golden files; deferred to a follow-up.

std.parseJson error prefix: Originally included in this PR as a
change from "Invalid JSON: " to "failed to parse json: ", but
reverted on follow-up review because go-jsonnet's full message
structure ("failed to parse JSON: <Go parser detail>") is not
reproducible from ujson's message format. Changing the prefix alone
produced a third, less-informative wording; left as "Invalid JSON: ".

Related work

These are pure error-message text fixes, orthogonal to the existing
behavioral fixes in PR #1004 (std.round), PR #1005 (escapeStringJson),
and PR #1006 (manifestPython float format).

Test plan

./mill sjsonnet.jvm.3_3_7.test passes (FileTests.new_test_suite + FileTests.go_test_suite + FileTests.test_suite + StdMathTests + TailCallOptimizationTests + EvaluatorTests + all other suites)
./mill sjsonnet.jvm.2_12_21.test passes
./mill sjsonnet.jvm.2_13_18.test passes
./mill sjsonnet.jvm.3_3_7.reformat reports "Everything is formatted already"
16 golden files regenerated; 3 test assertions updated
Verified against go-jsonnet 0.22.0, C++ jsonnet 0.22.0, jrsonnet 0.5.0-pre99
No behavior changes — only error-message text

Motivation: A number of sjsonnet error messages diverged from go-jsonnet's exact wording in casing or punctuation. Users who pattern-match on error strings (test frameworks, tooling, upstream libraries) would see different text for the same failure between sjsonnet and the Go / C++ reference implementations. Modification: - std.pow, std.asin, std.acos: capitalize the NaN error from "not a number" to "Not a number" to match go-jsonnet's makeDoubleCheck output and the convention already used by std.sqrt, std.log, std.log2, and std.log10 in sjsonnet (MathModule.scala). - Val.Num.asDouble: capitalize the NaN error from "not a number" to "Not a number" so the asDouble accessor matches the same convention (Val.scala). - std.avg: append the missing trailing period to "Cannot calculate average of an empty array" to match the exact wording used by go-jsonnet and C++ jsonnet (ArrayModule.scala). - std.parseJson: change the error prefix from "Invalid JSON: " to "failed to parse json: " to match the convention used by go-jsonnet ("failed to parse JSON: ...") and jrsonnet ("failed to parse json: ..."); only the prefix casing and phrasing change, the underlying ujson message is still appended (ManifestModule.scala). - Regenerated six golden files to reflect the new wording: error.math_pow_nan / asin_nan / acos_nan (new_test_suite), pow4 (go_test_suite), error.parse_json and error.std_parseJson.nodigitsep (test_suite). Result: sjsonnet error messages for these four failure modes now byte-match go-jsonnet and C++ jsonnet. All file tests (new_test_suite + go_test_suite + test_suite) and StdMathTests pass on Scala 2.12 / 2.13 / 3.3. Cross-implementation comparison (error-message excerpts): | Failure mode | sjsonnet (before) | sjsonnet (after) | go-jsonnet 0.22.0 | jrsonnet 0.5.0-pre99 | C++ jsonnet 0.22.0 | |---------------------------|-----------------------------------------|-----------------------------------------|-----------------------------------------|-----------------------------|------------------------------------------| | std.pow(-1, 0.5) NaN | "not a number" | "Not a number" | "Not a number" | "non-finite" | "not a number" | | std.asin(2) NaN | "not a number" | "Not a number" | "Not a number" | "non-finite" | "not a number" | | std.acos(2) NaN | "not a number" | "Not a number" | "Not a number" | "non-finite" | "not a number" | | std.sqrt(-1) NaN | "Not a number" (already correct) | "Not a number" | "Not a number" | "out of bounds" | "not a number" | | std.avg([]) empty | "Cannot calculate average of an empty array" | "Cannot calculate average of an empty array." | "Cannot calculate average of an empty array." | "expected non-empty array" | "Cannot calculate average of an empty array." | | std.parseJson("{bad") | "Invalid JSON: ..." | "failed to parse json: ..." | "failed to parse JSON: ..." | "failed to parse json: ..." | "[json.exception.parse_error.101] ..." | Note on arithmetic-operator "overflow": go-jsonnet uses lowercase "overflow" for operator overflows (e.g. 1.8e308 * 2) and uppercase "Overflow" for builtin overflows (e.g. std.exp(1000)). sjsonnet routes both paths through Val.Num's infinity check, so it cannot distinguish the two. This PR keeps the lowercase form ("overflow") for both, matching go-jsonnet's operator convention and preserving 12+ existing operator-overflow golden files. Builtin overflows (std.exp, std.log, std.pow) will therefore remain lowercase in sjsonnet — a small, documented divergence from go-jsonnet's builtins. Note on std.assertEqual: the message lives in stdlib.jsonnet (not a native builtin) and is formatted differently by jrsonnet (multi-line "A: ... / B: ..." form). Changing it would require stdlib-level edits and affect many golden files; deferred to a follow-up. ## Test plan - [x] `./mill sjsonnet.jvm.3_3_7.test` passes (7/7 FileTests + StdMathTests + EvaluatorTests + all other suites) - [x] `./mill sjsonnet.jvm.2_12_21.test` passes (38/38 tests) - [x] `./mill sjsonnet.jvm.2_13_18.test` passes (74/74 tests) - [x] `./mill sjsonnet.jvm.3_3_7.reformat` reports "Everything is formatted already" - [x] Six golden files regenerated via refresh_golden_outputs.sh-style invocation - [x] Verified against go-jsonnet 0.22.0, C++ jsonnet 0.22.0, jrsonnet 0.5.0-pre99

The "failed to parse json: " prefix introduced in the parent commit read less clearly than the original "Invalid JSON: " prefix. Since go-jsonnet's full message structure ("failed to parse JSON: <Go parser detail>") is not directly reproducible from ujson's message format, changing the prefix alone did not actually align with go-jsonnet — it just produced a third, less-informative wording. Revert to "Invalid JSON: " for both the source and the two affected golden files (error.parse_json.jsonnet.golden and error.std_parseJson.nodigitsep.jsonnet.golden in test_suite/).

…l numeric errors The prior PR capitalized NaN/Overflow messages for builtins (std.pow, std.asin, std.acos, Val.Num.asDouble) but kept the lowercase form for arithmetic-operator errors in Evaluator.scala and for Val.Num's constructor (used by any operation that yields Infinity). This created an inconsistency: std.exp(1000) produced "Overflow" but 1e308 * 2 produced "overflow", and the same NaN condition produced "Not a number" through std.sqrt but "not a number" through 0.0 / 0.0. User feedback was clear: pick one convention and apply it consistently. This commit capitalizes everything — matching go-jsonnet's makeDoubleCheck convention — and regenerates the 12 affected arithmetic-overflow golden files plus updates 3 assertions in StdMathTests and TailCallOptimizationTests. Changes: - Val.scala: Val.Num constructor "overflow" → "Overflow". - Evaluator.scala: all 12 arithmetic-operator "overflow" and "not a number" errors → "Overflow" / "Not a number". - 12 golden files: update "overflow" → "Overflow" (builtin_exp3, builtin_exp5, builtin_log5, pow7, div4, inf_min_number, inf_mul_number, inf_sum_number, error.arithmetic_overflow_*, error.overflow2). - StdMathTests.scala: log(0) / log2(0) assertions "overflow" → "Overflow". - TailCallOptimizationTests.scala: tailstrictFactorialOverflow assertion "overflow" → "Overflow". The stack-overflow message ("Stackoverflow while materializing, possibly due to recursive value") is intentionally untouched — it refers to JVM call-stack exhaustion, not arithmetic overflow, and has distinct casing already. Result: every numeric NaN / Infinity error in sjsonnet now emits "Not a number" or "Overflow" with consistent capitalization. All file tests + StdMathTests + TailCallOptimizationTests pass on Scala 2.12 / 2.13 / 3.3. Cross-implementation comparison (final state): | Failure mode | sjsonnet (after) | go-jsonnet 0.22.0 | jrsonnet 0.5.0-pre99 | C++ jsonnet 0.22.0 | |----------------------------|------------------|--------------------|----------------------|--------------------| | std.pow(-1, 0.5) NaN | "Not a number" | "Not a number" | "non-finite" | "not a number" | | std.sqrt(-1) NaN | "Not a number" | "Not a number" | "out of bounds" | "not a number" | | 0.0 / 0.0 NaN (operator) | "Not a number" | "Not a number" | "non-finite" | "not a number" | | std.exp(1000) Overflow | "Overflow" | "Overflow" | "non-finite" | "overflow" | | 1e308 * 2 Overflow (op) | "Overflow" | "overflow" * | "non-finite" | "overflow" * | * go-jsonnet / C++ jsonnet keep lowercase "overflow" for operator- level overflows. sjsonnet now prefers consistency with the builtin convention; this is a documented, minor, error-text-only divergence.

He-Pin · 2026-06-20T08:38:48Z

Superseded by #1010 which contains all 4 commits (the original PR #1007 + 3 follow-ups including the comprehensive capitalization sweep).

He-Pin added 3 commits June 20, 2026 13:03

He-Pin mentioned this pull request Jun 20, 2026

fix: unify and improve error-message quality across sjsonnet #1010

Open

He-Pin closed this Jun 20, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fix: align error-message casing and punctuation with go-jsonnet#1007

fix: align error-message casing and punctuation with go-jsonnet#1007
He-Pin wants to merge 3 commits into
databricks:masterfrom
He-Pin:worktree-fix-error-messages-batch

He-Pin commented Jun 20, 2026 •

edited

Loading

Uh oh!

He-Pin commented Jun 20, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

He-Pin commented Jun 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Modification

Result

Cross-implementation comparison (final state)

Intentionally deferred

Related work

Test plan

Uh oh!

He-Pin commented Jun 20, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

He-Pin commented Jun 20, 2026 •

edited

Loading