HermitCrab analysis: one-template battery is 45% of work on a private grammar; FST-derived precheck ceiling unmeasured

未關閉
  • 哪裡不對？
#604 0 則留言 0 個 reaction 已指派 0 人 在 GitHub 檢視

維護者通常 1 天內回覆

還沒有人認領這個 Issue。

評估

難度
4/5
預估耗時
3-5 天
新手友好度
38/100
Issue 類型
功能
描述清晰度
基本清楚
活躍度
活躍
技術堆疊
csharp
領域
performance

研究方向

Start with docs/hermitcrab-2026-10-21-search-cost-round.md, lines 319–330, on the docs/hc-optimization-ledger branch. Instrument zero-match traversal work in steps on single-template grammars and measure its share of parser work; the issue says to build a precheck only if that share is material.

由索引模型根據 Issue 內容生成。

描述

Claude-drafted.

On a one-template grammar the template battery is 45% of parser work (a private grammar variant, v1, held list; HC-Rust work steps, before change C), and the shipped shared-slot-batch change cannot help it. An FST-derived precheck could reject slot rules before traversal, but its ceiling is not measured.

Status: open research lead, not built.

What we know (HC-Rust port, work steps; the lead is tagged as having a C# analogue):

  • Battery share of parser work before change C: v1 held 45%, Mbugwe 61%, Sena 23%, Amharic 21%, private base 8%, private deletion variant 6.5%, Indonesian 0 (no templates).
  • Sharing trailing slot batches (change C, ported as #546) saves 0 work on Amharic and every private variant (one template each) and on Indonesian (no templates), so there is nothing to share.
  • After fix A, v1's held list still costs 105.8M work units, 45% of it in the one-template battery.
  • The nearest shipped check is the edge-segment prefilter (ledger row 22). It tests only a rule's first and last segments and was judged on traversals. This lead derives its condition from the whole compiled matcher and is judged in work steps, since a traversal charges per arc.

Not known: the ceiling, meaning the work spent in zero-match traversals on single-template grammars.

Proposed change: instrument zero-match traversal work in steps on single-template grammars first. Build the precheck only if that share is material.

Related: #564 (open), #574 (PR, open), #599 (open), #546 (PR, open). PanGloss: sillsdev/PanGloss#106 (pointer, open). The required-slot look-ahead in sillsdev/PanGloss#25 and #66 is a different mechanism (state not checked).

Next: add a zero-match traversal work counter and report it per single-template grammar.

Evidence -- battery share by grammar

Battery share of all parser work, HC-Rust, before change C (L191-L201 (lines 191–201)):

Grammar Battery share
Mbugwe 61%
Private v1, held 45%
Sena 23%
Amharic 21%
Private base 8%
Private, loaded from C#'s export 9%
Private deletion variant 6.5%
Indonesian 0

Change C's static census found no sharing possible on Amharic (50 to 50) and the private project (30 to 30), one template each (L204-L207 (lines 204–207)). Change C measured 0 on the grammars with one template (L222-L223 (lines 222–223)).

Sources -- pinned to 2eba2f50 (closed PR #490 record)
主要語言
C#
星號
29
分支
16
平均合併
4 天 11 小時
30 天內合併 PR
14

環境準備

  • 沒有 Dockerfile 或 Docker Compose 檔案
  • 有 Pull Request 範本
  • 沒有貢獻指南

從這裡開始

  1. 先讀完整個 Issue，再讀專案的貢獻指南。
  2. 在 Issue 下留言說明你要接手 —— 這能避免兩個人做同樣的事。
  3. Fork 儲存庫，在一個分支上完成修改。
  4. 送出 Pull Request，並在描述裡引用這個 Issue 編號。

sillsdev/machine 的其他 Issue

查看 sillsdev/machine 的全部 Issue

相似的 Issue

更多 C# Issue

把新 issue 寄到你的電子郵件信箱

精選適合新手參與的 GitHub issue 摘要。