test: consolidate 192 test files into 15 - #49
Merged
Merged
Conversation
Merges block_alloc_regression, codec_cap_regression, differential_ocf, example and large_datum into one file. package ocf_test cannot merge with package ocf, so this bucket has its own file by construction. Move only: every declaration is byte-identical, imports are the deduped union, and each section keeps a `// ---------- orig_test.go ----------` banner (the convention conformance_test.go already uses).
Framing, codec ownership and nil-spelling, decompression limits, the DoS battery, foreign-writer differentials, fuzz targets, the OCF matrices, metadata compat and the truncation sweeps now sit beside ocf_test.go. Move only: every declaration body is byte-identical to its source, the import block is the deduped union, and each section keeps a `// ---------- orig_test.go ----------` banner.
…_test.go The JVM differentials (schema, matrix, value, attribute-placement census) share one build constraint and one package; //go:build cisuite means they can never merge with anything outside it, so this bucket is one file by construction. Move only. Verified against the -tags=cisuite name list as well as the default one, since a plain `go test` never compiles these.
json_decode_test.go joins json_codec_test.go: both drive the Avro-JSON wire (json_codec.go / json_decode.go / json_scan.go) from inside package avro, and the encode and decode halves are read together. Move only.
…t.go The package-avro tests that are not about one subsystem but about driving the internals hard and measuring the result: the predicate-agreement and hidden-state censuses, the DoS entry-point battery, the race wall-clock authority, every fuzz target, the benchmarks, and the coverage-gap pins. export_test.go moves here too, as the `// ---------- export_test.go ----------` section. It is the bridge package avro_test reads (RaceRelaxedForTest / RaceEnabledForTest / the slab-free probes), and race_bounds_test.go, the authority it bridges, is now three sections above it. Worth knowing that the stdlib name is gone: grep the banner, not the filename. Move only.
The encode side of the binary wire: the array/map element fast paths, the zero-byte-item producer and unsafe compliance pins, empty-bytes identity, the per-element minimum stand-in, magnitude arithmetic, the union tag tiers, the TextAppender/TextMarshaler precedence contracts, and the decode-then-reencode symmetry audit. Move only.
The decode side and the machinery it reaches: schema resolution, map set-iteration targets, union branch matching, the integer-overflow and enum-ordinal-overflow audits, the unsafe depth bound, the logical gate, the json.Unmarshal-input integration net, and the CustomType, callback contract and hidden-state suites (custom wiring is decoder-side state). Move only.
Go type -> schema inference in one place: namespace-scope composition of a CustomType.Schema, stray structural keys, the generative round-trip and struct-shape nets, the tag-edge net, reserved-key exact case, the null spelling, and the whole embedding family (placement, selection, diamond cost, repeated embeds). These stay in package avro because the shape nets reach collectFields and typeFieldMapping; none of them can move to package avro_test. Move only.
Everything from schema text to the parsed tree and back: the SchemaNode metadata API and its namespace handling, the parse front-end diff, node kind compat, the name-lookup tables and reference-schema pins (#42), the caller-composed and caller-edited node matrices, the tree value-type matrix, alias resolution, fingerprint purity, single-object encoding, the SchemaCache suites, and the structural cost work (breadth, DAG cost, the budgeted walk census, walk-budget marshalling, depth uniformity, degenerate cardinality, union forward references). Move only.
…e_test.go conformance_test.go was already a merge of sixteen files under the `// ---------- orig_test.go ----------` convention; this extends it with the rest of the black-box, oracle-driven work: every fastavro differential (binary, JSON, logical, alias resolution, null spelling, matrix, SchemaFor scope), the Apache schema-tests.txt vectors, the property and invariant nets, the tier-2 boundary / cross-path / resolution / metadata / SchemaFor-round-trip matrices, the decimal round-trip matrix, and the nesting suite. example_test.go merges in as well. Its Example funcs are godoc examples and go/doc finds them by function name, not by filename, so pkg.go.dev is unaffected -- but the conventional filename is gone, so: they are the `// ---------- example_test.go ----------` section, or `grep -n "^func Example" conformance_test.go`. Move only.
The 30 matrix_* cell files plus the magnitude-arithmetic matrix and the CustomType resurrection parity battery now sit in the file that already owns the shared tables. Nothing is replaced: matFrags, matCtxs, matEqual and runCore stay exactly as they were, and the cells that read them (matEqual alone had 16 readers) are now in the same file as the definition. customFrags and resurrectionCells likewise -- resurrectionCells is read by matrix_generative_test.go, which is why the parity battery belongs in this group rather than with the regressions. Move only.
New file. The external suites that answer "what does this piece of schema text MEAN" as a class rather than as remembered cases: the attribute x placement acceptance census, the reserved-attribute enumeration and its duplicate / exact-case rules, stray precision-scale and stray-key routing, attributes written as their destination zero, type-level default and order binding, non-string logicalType values, the null spellings (quoted name, body presence, branch spelling), namespace and empty-name-component grammar, the enum default token, the union default-order / fixed-size / tag-table rules, forward-referenced and negative-zero defaults, the flat goavro-style field format, the struct-tag doc contract and runtime tag grammar, one-way text inference, the JSON escape-position and skip strictness pins, the reader-grammar census, encode error identity, the DoS-cap producer-compliance table, and the error-message bound. Move only.
…t.go The remaining black-box regressions, grouped with the audit battery they belong to: the CustomType family (named Avro types, skip-decode parity, record-level Decode, pointer-returning customs at union branches), the internal-reparse lax-name surfaces and their class matrix, the SchemaCache self-containment pins (canonical, overlapping splice defs, wildcard registration), the slab-free no-alloc oracle (#41), cyclic field types, named byte elements, the unsafe null-union nil peel, and the b-soft omitzero contract. Move only. This is the last of the 13 consolidation groups: 192 test files -> 15.
The suite reads its own source in several places, and consolidation moved
what those reads name. Repaired, not relaxed:
* TestInvariant_EveryBatteryEntryPointHasABreadthCell and
TestInvariant_DagCostMatrixDrivesEveryEntryPoint each read ONE file.
They now read that file's SECTION of the merged file, via a new
testFileSection helper keyed on the `// ---------- x ----------` banner.
Reading the whole enclosing file would be a false pass: the DAG-cost
guard looks for "cache.Parse(", which its own section uses twice and
the rest of schema_test.go uses fifty-six more times.
testFileSection terminates only on a banner naming a *_test.go file.
The same shape is also used for ordinary subsection headings ("the
shapes", "triggers", "the WIDTH axis"), and stopping at one of those
truncates a section to a few lines.
* moduleTestFiles required 25 test files; there are 15 now, so the floor
is 10. The directory half is unchanged and is the half that catches
the failure it was written for.
* raceRelaxations rows by file, and two rows are now one file
(race_bounds + export -> internal_nets_test.go, 3+1 sites) with a
third renamed (error_bound -> schema_semantics_test.go). Counts and
reasons carry over intact; the row covering two sections says so.
* blankCode now blanks RUNE literals. This is a latent bug the
consolidation exposed rather than caused: `s[0] != {rune-brace}`
contributes an unmatched brace, so the enclosing function extent runs
to EOF, and a rune-quoted double quote opens a phantom string that
blanks the code after it. One test measured 7249 lines instead of 97
and swallowed a cost generator declared far below it, producing a
false edge in TestInvariant_EveryCostCellDrivesItsFactor. Small files
hid it because the over-run hit EOF within a few hundred lines. It
fails silently in the other direction too, since a phantom string can
blank a genuine call.
Neuter-verified after the change: remove the only Fingerprint breadth
cell, add an unrowed race-predicate consult, add an unrowed test driving
a cost generator. All go red (exit!=0, RUN>0, no panic).
Pre-existing, found while repointing the guard at its section and proven by neuter, not by reading: renaming all six real "ocf-header" cells in the DAG-cost matrix left the guard GREEN. The want list lives inside the very file the guard reads, so each entry matched its own list line. The check could not go red for any edit to the matrix. The same is true at the original commit — the guard read schema_dag_cost_test.go, which is where its want list was — so this is not fallout from the consolidation. One entry, "Parse(dag)", occurs exactly once in that file, meaning the ONLY occurrence was its own list line: it asserted nothing from the day it was written. The matrix spells that cell wantCostDoesNotScale(t, cell, "Parse", ...). Two changes: the guard cuts its own function body out of the searched text (via t.Name(), so it cannot rot on a rename), and the wants become the cells LABELS, which is the naming the matrix actually uses and applies uniformly across all nineteen of its cells. Neuter-verified: renaming the six real ocf-header cells while leaving the want list intact — the exact edit that used to pass — now goes red (exit!=0, RUN>0, no panic).
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Moves test code between files. Deletes no test. Stacked on #48.
192 test files -> 15. The long tail was the problem, not the big files: 132 of
192 were under 500 lines, 60 under 200.
The floor is 7, and it is forced
Seven buckets can never merge — incompatible build tags and the internal /
external package split:
Gate
The set of test function names before and after is byte-identical in all
three build configurations:
Merging was done by an AST tool that asserts each source's post-import body
appears byte-for-byte in the result before deleting the source — passed for all
190 merged files. Sections carry
// ---------- orig_test.go ----------banners, the convention
conformance_test.goalready used.Shared matrix tables (
matFrags,matCtxs,matEqual,runCore,customFrags,resurrectionCells) were reused, never replaced. Nothing movedbetween packages, so no internal test lost unexported access.
gofmtclean,go vetclean, suite green,-racegreen.Two pre-existing defects it surfaced
The suite reads its own source in a few guards, so consolidation broke them.
Repairing them turned up two bugs that predate this work:
blankCodenever blanked rune literals.s[0] != '{'contributes anunmatched brace and
s[1] != '"'opens a phantom string. One test'smeasured extent came out 7,249 lines instead of 97, swallowing a cost
generator declared far below it. Small files hid it — the over-run hit EOF
within a few hundred lines. It fails silently in the other direction too,
since a phantom string can blank a genuine call.
TestInvariant_DagCostMatrixDrivesEveryEntryPointcould not fail. Itswant-list literals sit inside the very file it reads, so every entry matched
its own list line. Renaming all six real
ocf-headercells left it green.One entry,
"Parse(dag)", occurred exactly once in the file — its own listline — so it asserted nothing from the day it was written. Now excises its
own function from the searched body, reads the section rather than the file,
and is neuter-verified red.
To be clear about severity on (2): the entry points themselves were always
covered by
TestInvariant_EveryMinBytesEntryPointIsBounded. What had rottedwas the tripwire guarding the matrix against losing a cell, not the DoS bounds.