perf: close the v0.7 regressions on JSON models, XRPC writes, Jetstream and CAR parsing #190

New issue

Closed

opened 2026-09-27 21:26:47 +00:00 by Grandiras · 0 comments

Grandiras commented

2026-09-27 21:26:47 +00:00

Owner

Part of the v0.7.0 update cycle (tracking #141). The pre-release benchmark compares v0.6.0 (7f672c0) with main (e7f9e61). It uses BenchmarkDotNet 0.15.8 on .NET 10.0.12, AMD Ryzen 5 7600X, in-process. The DAG-CBOR, LtHash, MST and identifier paths got much faster, but five paths regressed:

Benchmark	v0.6.0	v0.7.0	time	alloc
`JsonBenchmarks.DeserializeGetTimelineResponse` (50 posts)	133.4 µs / 151.6 KB	186.5 µs / 213.1 KB	1.40×	1.41×
`XrpcBenchmarks.XrpcCreateRecord`	2.18 µs / 4.57 KB	2.75 µs / 4.71 KB	1.26×	1.03×
`JetstreamBenchmarks.ParseJetstreamCommitEvent`	1.40 µs / 2.02 KB	1.71 µs / 1.56 KB	1.22×	0.77×
`JsonBenchmarks.SerializePostRecord`	403 ns / 944 B	468 ns / 960 B	1.16×	1.02×
`CarBenchmarks.ParseSynthetic1000BlockCar` / `ParseCommitCar`	429 / 14.7 µs	437 / 15.6 µs	1.02–1.06×	1.17–1.22×

These most likely come from this cycle's correctness features:

typed-identifier and AtDatetime converters, which validate every value;
the open-union converter, which scans ahead for $type;
[JsonExtensionData] on every model;
the stricter CAR/CID checks.

Those features stay. Their cost has to come down.

Goal: each regressed benchmark is at or better than v0.6.0, or within ~5% with the reason documented, and none of the improved benchmarks gets worse. Behaviour must not change: the drift, convention and interop tests all pass.

Part of the v0.7.0 update cycle (tracking #141). The pre-release benchmark compares v0.6.0 (`7f672c0`) with main (`e7f9e61`). It uses BenchmarkDotNet 0.15.8 on .NET 10.0.12, AMD Ryzen 5 7600X, in-process. The DAG-CBOR, LtHash, MST and identifier paths got much faster, but five paths regressed: | Benchmark | v0.6.0 | v0.7.0 | time | alloc | |---|---:|---:|---:|---:| | `JsonBenchmarks.DeserializeGetTimelineResponse` (50 posts) | 133.4 µs / 151.6 KB | 186.5 µs / 213.1 KB | **1.40×** | **1.41×** | | `XrpcBenchmarks.XrpcCreateRecord` | 2.18 µs / 4.57 KB | 2.75 µs / 4.71 KB | 1.26× | 1.03× | | `JetstreamBenchmarks.ParseJetstreamCommitEvent` | 1.40 µs / 2.02 KB | 1.71 µs / 1.56 KB | 1.22× | 0.77× | | `JsonBenchmarks.SerializePostRecord` | 403 ns / 944 B | 468 ns / 960 B | 1.16× | 1.02× | | `CarBenchmarks.ParseSynthetic1000BlockCar` / `ParseCommitCar` | 429 / 14.7 µs | 437 / 15.6 µs | 1.02–1.06× | **1.17–1.22×** | These most likely come from this cycle's correctness features: - typed-identifier and `AtDatetime` converters, which validate every value; - the open-union converter, which scans ahead for `$type`; - `[JsonExtensionData]` on every model; - the stricter CAR/CID checks. Those features stay. Their cost has to come down. **Goal:** each regressed benchmark is at or better than v0.6.0, or within ~5% with the reason documented, and none of the improved benchmarks gets worse. Behaviour must not change: the drift, convention and interop tests all pass.

Grandiras referenced this issue

2026-09-27 21:27:32 +00:00

refactor!: pre-release debloat pass #180

Grandiras referenced this issue from a commit

2026-09-27 23:14:39 +00:00

perf: close the v0.7 regressions on JSON models, XRPC writes, Jetstream and CAR parsing (closes #190)

Grandiras referenced this issue from a pull request that will close it,

2026-09-27 23:14:40 +00:00

perf: close the v0.7 regressions on JSON models, XRPC writes, Jetstream and CAR parsing #191

Grandiras referenced this issue

2026-09-27 23:14:49 +00:00

perf: close the v0.7 regressions on JSON models, XRPC writes, Jetstream and CAR parsing #191