Repository navigation
Conversation
Contributor
Test ReportTest Results
Code Coverage (Java 25)
Changed Class Coverage (11 classes)
|
bbakerman
approved these changes
Nov 4, 2026
…generation-performance
Member
Author
|
@xuorig could you give this a go before merging and see if this improves things for you? |
Contributor
|
Hi @andimarek sorry I had missed this thread. I'll give this a try tomorrow. |
Contributor
|
Hi @andimarek, I won't have time to test this this week due to how our fork is setup, need to port it carefully. But I can tell you the benefits of this will likely be huge. 🚢 |
This branch has not been deployed
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
This PR reduces SDL checking, directive construction and shared traversal work during
schema generation. The direct IDL follow-up below affects both
SchemaGeneratorand
FastSchemaGeneratorthrough the checker and construction helpers they call.Direct SDL checking and construction changes
SchemaTypeChecker.checkFieldTypesPresentSchemaTypeChecker.checkNamedUniquenessSet.addto identify duplicates.ImplementingTypesCheckerSchemaTypeDirectivesCheckerSchemaGeneratorHelper.BuildContextSchemaGeneratorAppliedDirectiveHelperTypeDefinitionRegistryis unchanged in this follow-up: the existing registryimprovements are already in the baseline, and the profiles supported the checker
and directive-construction changes above.
Final incremental mean time reduction versus follow-up baseline
baf539554(the previous PR plus master
96443c271; the same master correctness fixes arepresent in both variants). Positive means less time; negative means more:
Allocation per operation is lower in all 20 cases, by 1.62-61.84%. Time means
are lower in 19/20. SchemaGenerator parsing fixture 3 is 0.23% slower, with
overlapping fork means and JMH intervals; no time gain is claimed for that case.
Timing intervals also overlap for SchemaGenerator fixture 4 in both scopes and
all fixture 5 cases. Those small gains are observed means, not separately
established significant improvements.
The two change groups were measured independently before the final comparison:
These one-fork comparisons attribute results to groups, not individual edits,
and do not enter the final gains above. All ten Fast cases had lower time means
in the directive experiment. Unfavorable exploratory cases are retained: notably A's
SchemaGenerator fixture 5 slowdowns and B2's fixture 4/5 allocation increases
versus A. Separate profiles and repeated forks showed allocation variation in
unchanged traversal paths in both A and B2; JVM allocation elimination is a
possible explanation, not proven causation. The final combined comparison has
lower mean time/allocation in those cases, but does not independently disprove
every unfavorable A-versus-B2 difference.
Methodology: dedicated Ryzen 5 3600; Java 25 only, 24 GiB fixed heap, ZGC,
compact object headers, adaptive GC workers. All five fixtures, both generators,
both scopes: 20 timing cases, three forks/version/case, each with 10 x 10s warmup
and 5 x 10s measurement. Version order alternates by case; runs are serial under
a global lock. Nineteen separate async-profiler recordings used the same warmup
and 3 x 10s measurement; their timing scores are excluded. All 120 final timing
forks completed with zero allocation stalls. Two candidate warmup trends were
inspected and retained (SG generation fixture 4 and parsing fixture 5, fork 3);
their uncertainty is documented. Original governors and paused timer/service
states were restored and verified after measurement.
Correctness: 529 targeted IDL cases; full Java 25 ZGC/compact-header suite
6,146 passed, seven existing skips; all ten generated-schema hashes/type counts
match baseline; Javadoc and the per-class coverage gate passed. Local totals
include two unrelated untracked parser tests excluded from this PR. Spock tests
cover error order, wrapper comparisons, extensions, malformed interfaces,
directive defaults/nulls/unset values, both generators and options, ordering,
forward definitions and build isolation. CI passed Java 17/21/25 tests and TestNG, jcstress, Javadoc, build checks, file validation, and the per-class coverage gate.
Complete follow-up results, both exploratory tables, raw JMH/GC logs, profiles, limitations and reproduction.
The follow-up measurements are incremental to a newer baseline. Keep them
separate from the original results below; do not add or multiply percentages.
Original shared traversal changes and results
Both schema generators spend substantial time traversing schema elements and validating the resulting schema. This PR reduces allocation and repeated work in that shared code through four changes, developed over three profiling rounds:
Create ancestor lists only when needed —
DefaultTraverserContext. Previously, every traversal context copied its parent's complete ancestor list, even when the visitor never requested that list. Now it captures the parent breadcrumb and builds/caches the full immutable list on the firstgetBreadcrumbs()call. Traversals that do not need breadcrumbs avoid those list allocations and copies. Parent-node snapshots and custom context behavior are preserved.Create local variable maps only on the first write —
TraverserState/DefaultTraverserContext. Previously, every non-root sequential traversal context allocated an emptyLinkedHashMap. Now the map is allocated only whensetVar()is called. Contexts that only read inherited state therefore need no local map; parent lookup and variable shadowing remain dynamic.Allocate the flattened child list at its final size —
SchemaElementChildrenContainer.getChildrenAsList(). Previously, repeatedaddAll()calls grew the result array and created temporary arrays for child groups. Now the method sums the child counts, allocates the required capacity once, and appends children directly. This removes temporary arrays and growth copies while preserving order, duplicates, and an independent mutable result.Look up validation state only for nodes with directives —
AppliedDirectivesAreValid. Previously, every visited node searched its ancestor contexts for the schema and error collector before checking whether it had directives. Now that check happens first. Nodes without directives avoid both ancestor searches; nodes with directives receive the same validation checks.The JMH benchmarks now cover both generators, all five large schemas, and both generation-only and parsing-plus-generation, with schema validation enabled. Spock tests cover the behavior affected by these optimizations.
Final mean time reduction versus master
b88bb44f5(already includes #4528):Allocation per operation falls 2.48–30.32%. The 2.17% Fast/fixture-2 parsing gain has overlapping variability; it is not claimed as independently significant.
Measured serially on a dedicated Ryzen 5 3600 using Java 25, ZGC, compact headers, adaptive GC workers, and a 24 GiB heap. Final comparison: 3 forks/version/case, each with 10 × 10s warmup and 5 × 10s measurement, alternating version order. Profiling was separate. All 120 forks completed without allocation stalls; one warmup trend was inspected and documented. Original server settings were restored.
Validation: 6,029 Spock cases passed (7 existing skips), all ten generator/fixture outputs match baseline, build/TCK checks passed, and the local per-class coverage gate reports no regressions. Local counts include two pre-existing parser tests excluded from this PR. CI also passed Java 17/21/25, jcstress, Javadoc, and the per-class coverage gate.
Full results, raw JMH samples, profile evidence, caveats, and reproduction instructions.