detectOAS32SchemaFeatures is called with a freshly allocated make(map[*parser.Schema]bool) at each site — once per media type, once per component schema. On a document with many media types that is one map allocation each, plus a full schema walk.
Measurements
BenchmarkConvertParsedOAS3ToOAS2, allocs/op:
|
v1.58.0 |
v1.59.0 |
| Small |
209 |
268 |
| Medium |
2208 |
2924 |
Attribution, by reverting individual walks:
| Source |
Small |
Medium |
mt.Schema walk (#426) |
+48 of 59 |
+109 of 716 |
| Rest of the 3.2 detection walk (#412) |
+11 |
+607 |
Why this is not filed as a regression
converter/oas32_features.go did not exist at v1.58.0 — it was created in #412. Before this release, oastools convert -t 2.0 reported nothing at all about OAS 3.2 fields. The cost buys output that did not previously exist, so there is no like-for-like baseline to regress against. Accepted for v1.59.0 and called out in the release notes.
What is worth fixing is the per-call map allocation, which is incidental rather than inherent.
The catch
Sharing one visited map across schemas is not behavior-preserving. Each call currently gets a fresh set, so a schema reachable from two paths is reported at each path. One shared map would report only the first, silently dropping findings.
So the safe shapes are:
- Allocate lazily, on first recursion rather than at the call — most inline schemas never recurse.
- Keep per-call sets but pool or reuse the allocation.
Option 1 mirrors what validator/schema_traversal.go does for its callback visited set, though note the measurement there: the equivalent change produced no measurable improvement, because the compiler was already eliding the allocation. Measure before claiming a win here too.
Acceptance
BenchmarkConvertParsedOAS3ToOAS2 allocs/op improves measurably, or the finding is closed with the measurement showing why it cannot- A schema reachable from two paths is still reported at both
detectOAS32SchemaFeaturesis called with a freshly allocatedmake(map[*parser.Schema]bool)at each site — once per media type, once per component schema. On a document with many media types that is one map allocation each, plus a full schema walk.Measurements
BenchmarkConvertParsedOAS3ToOAS2, allocs/op:Attribution, by reverting individual walks:
mt.Schemawalk (#426)Why this is not filed as a regression
converter/oas32_features.godid not exist at v1.58.0 — it was created in #412. Before this release,oastools convert -t 2.0reported nothing at all about OAS 3.2 fields. The cost buys output that did not previously exist, so there is no like-for-like baseline to regress against. Accepted for v1.59.0 and called out in the release notes.What is worth fixing is the per-call map allocation, which is incidental rather than inherent.
The catch
Sharing one visited map across schemas is not behavior-preserving. Each call currently gets a fresh set, so a schema reachable from two paths is reported at each path. One shared map would report only the first, silently dropping findings.
So the safe shapes are:
Option 1 mirrors what
validator/schema_traversal.godoes for its callback visited set, though note the measurement there: the equivalent change produced no measurable improvement, because the compiler was already eliding the allocation. Measure before claiming a win here too.Acceptance
BenchmarkConvertParsedOAS3ToOAS2allocs/op improves measurably, or the finding is closed with the measurement showing why it cannot