Experiment of rewriting in Ruby by yosiat · Pull Request #218 · yosiat/panko_serializer

yosiat · 2026-03-30T10:05:35Z

No description provided.

yosiat · 2026-03-30T10:42:41Z

Performance Analysis: Ruby Implementation vs C Extension

Benchmarks run on Ruby 4.0.2 (2026-03-17, +PRISM, +YJIT, arm64-darwin25), Rails 8.0.
Both master and ruby-impl-perf run with YJIT enabled (Ruby 4.0 default).

panko_json — JSON string output (primary benchmark)

Throughput (iterations/sec, higher is better)

Benchmark	Master (C ext)	Ruby impl	Change
Simple, 50 posts	38,566	54,443	+41%
Simple, 2300 posts	862	1,434	+66%
HasOne, 50 posts	21,771	32,777	+51%
HasOne, 2300 posts	500	842	+68%
HasMany, 50 authors	18,086	20,404	+13%
HasMany, 2300 authors	409	500	+22%
MethodCall, 50 posts	40,162	44,360	+10%
MethodCall, 2300 posts	937	1,128	+20%
JSON column, 50 posts	21,386	25,227	+18%
JSON column, 2300 posts	476	595	+25%
Except, 50 posts	23,269	35,347	+52%
Except, 2300 posts	541	892	+65%
Only, 50 posts	28,035	46,516	+66%
Only, 2300 posts	624	1,253	+101%
Aliases, 50 posts	36,130	51,859	+44%
Aliases, 2300 posts	847	1,364	+61%

Allocations (allocs / retained)

Benchmark	Master (C ext)	Ruby impl
Simple, 50	122 / 0	35 / 4
Simple, 2300	4,620 / 0	35 / 4
HasOne, 50	128 / 0	63 / 7
HasOne, 2300	4,628 / 0	62 / 6
HasMany, 50	558 / 100	438 / 106
HasMany, 2300	24,928 / 4,500	18,058 / 4,506
MethodCall, 50	22 / 0	37 / 2
MethodCall, 2300	22 / 0	37 / 2
JSON column, 50	170 / 0	86 / 5
JSON column, 2300	6,920 / 0	2,335 / 5
Except, 50	131 / 0	66 / 6
Except, 2300	4,631 / 0	66 / 6
Only, 50	31 / 0	66 / 4
Only, 2300	31 / 0	66 / 4
Aliases, 50	120 / 0	35 / 4
Aliases, 2300	4,620 / 0	35 / 4

panko_object — Ruby Hash/Array output

Benchmark	Master (C ext)	Ruby impl	Change
Simple, 50 posts	29,576	43,133	+46%
Simple, 2300 posts	627	1,031	+64%
HasOne, 50 posts	16,323	25,126	+54%
HasOne, 2300 posts	360	599	+66%
Except, 50 posts	17,247	25,951	+50%
Except, 2300 posts	389	626	+61%
Only, 50 posts	19,935	32,830	+65%
Only, 2300 posts	453	817	+80%
Aliases, 50 posts	29,659	42,098	+42%
Aliases, 2300 posts	628	1,032	+64%

plain_object — non-AR objects

Benchmark	Master (C ext)	Ruby impl	Change
Simple, 50 posts	43,619	43,107	−1%
Simple, 2300 posts	1,022	1,030	+1%
HasOne, 50 posts	32,185	30,938	−4%
HasOne, 2300 posts	752	724	−4%
MethodCall, 50 posts	61,258	44,904	−27%
MethodCall, 2300 posts	1,510	1,049	−31%
Except, 50 posts	33,668	32,577	−3%
Except, 2300 posts	829	797	−4%
Only, 50 posts	54,527	49,843	−9%
Only, 2300 posts	1,370	1,220	−11%

object_writer — Oj::StringWriter layer

Benchmark	Master (C ext)	Ruby impl	Change
1 prop, push_value	9,484,472	9,390,096	−1%
2 props, push_value	8,015,880	7,960,216	−1%
1 prop, push_key+push_value	8,988,532	9,030,661	0%
2 props, push_key+push_value	7,172,322	7,388,748	+3%
Nested object	5,665,602	5,711,066	+1%

type_casts — selected highlights

Type	Master (C ext)	Ruby impl	Change
String TypeCast	22,627,776	18,459,316	−18%
String NoTypeCast	56,552,541	31,026,084	−45%
Integer TypeCast	39,958,641	24,070,938	−40%
Integer NoTypeCast	56,278,885	32,497,737	−42%
Float TypeCast	29,732,308	20,437,117	−31%
Float NoTypeCast	50,937,768	33,383,268	−34%
Boolean TypeCast	40,481,193	5,393,576	−87%
Boolean NoTypeCast	55,342,118	32,013,417	−42%
DateTime TypeCast	24,132,573	10,609,659	−56%
DateTime NoTypeCast	10,735,715	9,903,270	−8%
Date TypeCast	2,755,819	2,672,488	−3%
Date NoTypeCast	14,129,258	12,562,597	−11%
Decimal TypeCast	3,050,253	2,977,445	−2%
Decimal NoTypeCast	6,254,275	6,405,401	+2%
Json TypeCast	3,243,300	3,248,083	0%
Json NoTypeCast	54,932,784	15,904,468	−71%

components — Ruby impl internal benchmarks (no C ext equivalent)

Component	ips
Filters no-op	1,172,151
Filters :only array	239,871
Filters :only hash (nested)	152,569
RecordState fast path (same batch)	9,935,600
RecordState full path (same class)	9,728,526
RecordState class change	66,959
Descriptor no filters	1,049,730
Descriptor attribute filter	399,191
Descriptor association filters	149,010

Summary

The Ruby implementation is faster than the C extension for all ActiveRecord serialization. On Ruby 4.0 with YJIT (the default), panko_json throughput improves +10% to +101% across every benchmark. The largest gains are on filtered queries: Only +101%, HasOne +68%, Except +65%, Simple +66% at scale. Even the smallest gain (MethodCall, 50 posts) is still +10%.

panko_object sees even larger gains (+42% to +80%). The Ruby Hash/Array output path benefits from the same YJIT optimizations. Only +80% on 2300 records is the standout.

plain_object is a regression. Non-AR serialization is −1% to −31% slower, with MethodCall taking the biggest hit (−27% to −31%). Simple attribute access is flat (±1%). This is the one area where the C extension still wins.

object_writer is unchanged (~0%). Expected — ObjectWriter is a pure Ruby stack-based Hash/Array builder that is identical in both branches.

Type cast microbenchmarks are slower in pure Ruby. Individual type conversions run 2–87% slower, with Boolean TypeCast (−87%) and Json NoTypeCast (−71%) being the outliers. However, these microbenchmarks measure isolated type conversions at millions of iterations/sec — the overhead is invisible in end-to-end serialization where the Ruby impl is faster overall.

Allocations are dramatically lower. The Ruby impl now allocates significantly fewer objects than the C extension in most benchmarks. For example, Simple 2300 drops from 4,620 to 35 allocations — a 99% reduction. Except 2300 drops from 4,631 to 66. HasMany 2300 drops from 24,928 to 18,058. The only cases where the Ruby impl allocates more are MethodCall (37 vs 22) and Only (66 vs 31), both small in absolute terms. This is a major improvement over the previous iteration where the Ruby impl allocated ~50% more objects.

rus-max · 2026-03-31T04:00:02Z

📊 Benchmarks: Ruby 4.0.1, Rails 8.1.0
Base: Master without YJIT = 100%

Benchmark	Master (no YJIT)	Master (YJIT)	PR (no YJIT)	PR (YJIT)
Simple 2300	100%	105% (+5%)	70% (−30%)	147% (+47%)
Simple 50	100%	107% (+7%)	68% (−32%)	133% (+33%)
HasOne 2300	100%	138% (+38%)	88% (−12%)	215% (+115%)
HasOne 50	100%	139% (+39%)	85% (−15%)	204% (+104%)

• Master + YJIT: 5-39% speedup
• PR without YJIT: 12-32% slower
• PR + YJIT: 33-115% faster 🚀

yosiat · 2026-03-31T18:04:38Z

@rus-max thanks for this comment.. I was pretty sure that the benchmark auto enable YJIT but this is the not the case.. I updated the performance comments and the results are impressive. Hopefully soon I'll be able to release the ruby version :)

rus-max · 2026-04-01T04:11:18Z

@yosiat Allocation benchmarks show no difference between Ruby impl and Ruby+YJIT.

yosiat · 2026-04-01T07:30:23Z

@rus-max I am not too familiar with YJIT internals, but I don't expect it to reduce allocations.
I am right now working on cleaning up all benchmarks files and simplify them, later I'll work on allocations, I think I have easy wins here.

…fter first call. +3.6% IPS Result: {"status":"keep","total_ips":13443.41,"total_allocs":4818}

…aseline! Simple 2300: 209 ips (was 175) Result: {"status":"keep","total_ips":15600.3,"total_allocs":4818}

…er baseline! Simple 2300: 265 ips Result: {"status":"keep","total_ips":18463.56,"total_allocs":4818}

…11% over baseline. 14422 vs 12978 Result: {"status":"keep","total_ips":14421.88,"total_allocs":4818}

…allocs unchanged at 4818 Result: {"status":"keep","total_ips":17167.58,"total_allocs":4818}

…6% over baseline Result: {"status":"keep","total_ips":17339.66,"total_allocs":4818}

…one call. +40.5% over baseline Result: {"status":"keep","total_ips":18243.27,"total_allocs":4818}

The 5-argument form of bytesplice was added in Ruby 3.3. Fall back to the 3-argument form with an explicit byteslice to support Ruby 3.2.

@Aliases

… C extension fallback Deletes the unused Context class from context.rb, the @Aliases attr_accessor and apply_fields_filters method from SerializationDescriptor, the stale aliases={} init in Serializer#inherited, and the C extension fallback branch in Impl::Serializer#write_fields that referenced non-existent Panko._sd_set_writer and Panko._write_attributes methods. Updates the one spec that asserted on aliases.

… cold paths

…-convention visibility

Move apply_filters, resolve_filters, apply_attribute_filters, and apply_association_filters out of SerializationDescriptor into a new Panko::Filters class backed by a single frozen instance (INSTANCE). SerializationDescriptor#apply_filters becomes a one-liner delegation. Adds spec/unit/panko/filters_spec.rb with full unit coverage.

…d edge-case specs

@type

- Remove @record_class ivar and all AR alias-resolution logic from Attribute - Simplify invalidate! to take no arguments (just clears @type and @cached_writer) - Move AR attribute_aliases resolution into ActiveRecord::Writer where it belongs - Expose alias_name= writer (attr_accessor) and make name= public - Add spec/unit/panko/attribute_spec.rb covering the new behaviour

Extract all record-level state (column_indexes, row, is_indexed_row, attributes_hash, types, additional_types, values, etc.) from the 252-line Writer god object into a new RecordState class. RecordState owns: - setup(object) — handles the IndexedRow fast path (identity check on column_indexes) and the full initialization path; returns true when the record class changes so the caller knows to invalidate attributes and re-resolve AR aliases - read_attribute(attribute) — non-indexed (Rails 7.x) value lookup with dirty-attributes-hash priority Writer becomes a thin orchestrator: calls record_state.setup, handles attribute invalidation and alias resolution on class change, then delegates into the existing indexed-row fast paths using state exposed via record_state accessors. Adds spec/unit/panko/impl/record_state_spec.rb with 29 unit examples covering initialization defaults, setup return values, fast-path detection, class-change tracking, and read_attribute priority logic.

… labels

…c conventions

…ngine The Impl namespace was a historical artifact from when this code was a direct port of the C extension. Now that it is idiomatic Ruby, Engine better reflects its role: the internal hot-path machinery that powers serialization, distinct from the public Panko::* API surface. - Rename lib/panko/impl/ → lib/panko/engine/ (git mv) - Rename spec/unit/panko/impl/ → spec/unit/panko/engine/ (git mv) - Replace all Panko::Impl → Panko::Engine references across lib/, spec/, benchmarks/, and panko_serializer.rb - Update CLAUDE.md to reflect new paths and namespace - Add .DS_Store to .gitignore

Replace 29 benchmark files with a clean 12-file structure: - support/benchmark.rb: ~200 LOC infra with benchmark() and benchmark_with_records() API, BENCH=/SIZE=/PROFILE= env vars, YJIT - support/setup.rb: SQLite in-memory DB with 2300 seed records - support/datasets.rb: 4 reusable datasets (posts, authors, aliased, plain) - 5 benchmark files: panko_json, panko_object, plain_object, object_writer, components (Filters, RecordState, SerializationDescriptor) - type_casts/: per-provider files (generic, postgresql, mysql, sqlite) Simplify Rakefile from PTY/JSON runner to simple system() calls. Remove active_model_serializers and terminal-table dependencies.

… allocations On Ruby 3.3+, bytesplice(dst_off, dst_len, src, src_off, src_len) copies directly without allocating intermediate byteslice strings. This removes ~6,900 object allocations and ~276KB per 2,300-record serialization run. Falls back to the 3-arg bytesplice + byteslice path on Ruby < 3.3.

association().target returns nil when the association hasn't been loaded yet. Check loaded? first and fall back to public_send for lazy loading.

Break the monolithic 180-line method into 9 small methods with clear responsibilities: write_attributes (dispatcher), handle_class_change, resolve_type, write_value, write_indexed_with_hash, write_indexed_cached, write_indexed_first_pass, write_non_indexed, and build_column_caches. Add nil_safe_push? to all value writers, replacing the is_a? chain in build_column_caches. Expose last_record_class on RecordState.

…. serialize) The subtype branch in ValuesWriter::Writer#write never cached a writer on the attribute, so the second record in a batch hit write_indexed_cached with a nil writer_cache entry and raised NoMethodError. Introduce SubtypeWriter to wrap the AR type and cache it like every other writer. Also removes stale TODO comments from attribute.rb and writer.rb.

- design-choices.md: Replace C extension references with Ruby Engine architecture, document fast paths, value writers, and IndexedRow - performance.md: Update benchmarks to Ruby 4.0.2/Rails 8.0, simplify to core JSON benchmarks, add instructions for running locally

Passing non-Array/Hash values to :only/:except now raises ArgumentError instead of NoMethodError, giving callers a clear message.

…imeWriter buffer corruption The global @@writer singleton shared a DateTimeWriter with a mutable @buf across all threads. Under Puma, concurrent requests could corrupt datetime output by writing to the same buffer simultaneously. Each thread now gets its own Writer instance via Thread.current.

…ociations When a serializer declares has_one pointing to a plain method on the model (not a real AR association), object.association() raises AssociationNotFoundError. Rescue and fall back to public_send, matching has_many behavior which already uses public_send directly.

@object

…ocal caching When Serializer.new is called without :only/:except/:context/:scope options, SerializationDescriptor.build now returns a thread-local cached copy instead of duplicating the descriptor on every call. This avoids allocating ~8 objects per call (descriptor, 2 array dups, serializer instance, 3 association duplicates with recursive sub-descriptors). Each thread gets its own duplicate (created once, reused forever), keeping the mutable per-call state (serializer @object, association Writer/RecordState) thread-safe under Puma. Also fixes a latent bug in RecordState#setup where the IndexedRow fast path could crash when a reused RecordState transitions from an IndexedRow-backed object to a non-indexed object. Benchmark (panko_oj, single Game object, Rails 8.0): Unpersisted: 119k → 190k i/s (+59%) Persisted: 96k → 171k i/s (+78%) Allocations per call: 79 → 23 objects (-71%)

Instead of creating a new Engine::Serializer on every serialize_to_json / to_json call, cache it on the descriptor via engine_serializer. Since descriptors are thread-local (from the previous commit), the cached engine is also per-thread and safe to reuse. Both write_fields and _serialize_many validate that the cached attributes_writer matches the current object type via AttributesWriter.writer_for, handling the edge case where a reused engine encounters a different object type (e.g. AR model then Hash). Also refactors AttributesWriter.create into writer_for (returns class) and create (instantiates), enabling cheap is_a? checks without allocation. Benchmark (panko_oj, single Game object, Rails 8.0): Unpersisted: 190k → 225k i/s (+18%) Persisted: 171k → 213k i/s (+25%) Cumulative from baseline: Unpersisted: 120k → 225k i/s (+88%) Persisted: 96k → 213k i/s (+122%) panko_json (array serialization): no change — one engine per batch.

yosiat force-pushed the ruby-impl-perf branch 3 times, most recently from 593c73d to 8be0e83 Compare March 30, 2026 20:51

yosiat force-pushed the ruby-impl-perf branch 12 times, most recently from d99c926 to 0d6f037 Compare April 5, 2026 11:30

yosiat force-pushed the ruby-impl-perf branch from 12c17e6 to 67c99b6 Compare April 10, 2026 09:11

yosiat added 9 commits April 11, 2026 16:11

Re-write in Ruby

29bb7ee

gitignore: add mise.toml and docs/superpowers

ef5ccab

Cache resolved writer on attribute object - skip case/when dispatch a…

ac5357d

…fter first call. +3.6% IPS Result: {"status":"keep","total_ips":13443.41,"total_allocs":4818}

Inline indexed row read in write_attributes hot path. +20% IPS over b…

ef59087

…aseline! Simple 2300: 209 ips (was 175) Result: {"status":"keep","total_ips":15600.3,"total_allocs":4818}

Move cached writer check before type.nil?/respond_to? checks. +42% ov…

2744941

…er baseline! Simple 2300: 265 ips Result: {"status":"keep","total_ips":18463.56,"total_allocs":4818}

Switch to Ruby datetime writer (remove C extension override). Still +…

d4654a6

…11% over baseline. 14422 vs 12978 Result: {"status":"keep","total_ips":14421.88,"total_allocs":4818}

bytesplice 5-arg form for datetime (zero-alloc). +32% over baseline, …

6b8bced

…allocs unchanged at 4818 Result: {"status":"keep","total_ips":17167.58,"total_allocs":4818}

Cache types/additional_types per model class in set_from_record. +33.…

e5b6a46

…6% over baseline Result: {"status":"keep","total_ips":17339.66,"total_allocs":4818}

Positional args in hot loop + cached assoc/method flags + direct has_…

868edaa

…one call. +40.5% over baseline Result: {"status":"keep","total_ips":18243.27,"total_allocs":4818}

yosiat added 28 commits April 11, 2026 16:11

Rubocop auto-correct

9191e66

Fix datetime bytesplice for Ruby 3.2 compatibility

7eec42e

The 5-argument form of bytesplice was added in Ruby 3.3. Fall back to the 3-argument form with an explicit byteslice to support Ruby 3.2.

Remove C extension — datetime conversion is now pure Ruby

b1e3e37

Consolidate serialize_one variants; replace while loops with .each in…

e45c9b6

… cold paths

Fix RDoc on _serialize_one: accurately describe callers and public-by…

de9ff11

…-convention visibility

Fix filter_associations to not mutate input arrays; add regression an…

4c47c4d

…d edge-case specs

Fix Writer initialize: declare all cache ivars; fix fast-path comment…

020fb42

… labels

Fix rubocop offenses in serializer, record_state_spec, and Gemfile

49394b0

Document rubocop as a hard requirement before every commit

415a9b7

Update CLAUDE.md: current architecture, architectural invariants, RDo…

c45199d

…c conventions

Fix has_one returning nil when association is not eager-loaded

3de90f6

association().target returns nil when the association hasn't been loaded yet. Check loaded? first and fall back to public_send for lazy loading.

Raise ArgumentError for invalid filter types and remove stale TODOs

2b45eb7

Passing non-Array/Hash values to :only/:except now raises ArgumentError instead of NoMethodError, giving callers a clear message.

Fix empty if branch in write_fields — use unless for clarity

60de606

Fix CLAUDE.md commands: all test/benchmark commands must use appraisal

36cb4a0

yosiat force-pushed the ruby-impl-perf branch from 8bbfc53 to 124a54a Compare April 11, 2026 13:11

yosiat mentioned this pull request Apr 12, 2026

Ruby impl perf codegen #226

Draft

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Experiment of rewriting in Ruby#218

Experiment of rewriting in Ruby#218
yosiat wants to merge 62 commits into
masterfrom
ruby-impl-perf

yosiat commented Mar 30, 2026

Uh oh!

yosiat commented Mar 30, 2026 •

edited

Loading

Uh oh!

rus-max commented Mar 31, 2026

Uh oh!

yosiat commented Mar 31, 2026

Uh oh!

rus-max commented Apr 1, 2026

Uh oh!

yosiat commented Apr 1, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

yosiat commented Mar 30, 2026

Uh oh!

yosiat commented Mar 30, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Performance Analysis: Ruby Implementation vs C Extension

panko_json — JSON string output (primary benchmark)

Throughput (iterations/sec, higher is better)

Allocations (allocs / retained)

panko_object — Ruby Hash/Array output

plain_object — non-AR objects

object_writer — Oj::StringWriter layer

type_casts — selected highlights

components — Ruby impl internal benchmarks (no C ext equivalent)

Summary

Uh oh!

rus-max commented Mar 31, 2026

Uh oh!

yosiat commented Mar 31, 2026

Uh oh!

rus-max commented Apr 1, 2026

Uh oh!

yosiat commented Apr 1, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

yosiat commented Mar 30, 2026 •

edited

Loading