Remove reachability-based memory sweeps from PerlOnJava

未關閉
#2,043 7 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
5/5
預估耗時
一週以上
新手友好度
5/100
Issue 類型
重構
描述清晰度
需要釐清
活躍度
活躍
技術堆疊
java, perl
領域
compilers

研究方向

Start by inspecting the WIP candidate branch wip/issue2043-preserve-diagnostic-20261113-082631 at commit 386856ab30, not the older PR #1938 head. Trace the failing DBIx phantom-chain regression from its acquisition point through the owner-slot lifetime boundary, using the existing diagnostic trace only to locate the mismatch. The integrator owns shared owner-ledger and lifecycle files; the sweep/walker removal and system-Perl evidence belong in separate worktrees. Done means no graph walk or sweep changes liveness, and focused JVM/interpreter checks plus t/52leaks.t, t/100populate.t, and CPAN acceptance pass.

由索引模型根據 Issue 內容生成。

描述

area:backend area:memory area:runtime enhancement

Goal

Remove reachability-based lifecycle decisions and memory sweeps. Perl strong references must use explicit owner slots with balanced acquisition and release. Historical origin: #1649.

Current checkpoint — 2026-11-19

PR #1938 remains open at remote head 75dd7d542f1f86bff930beb1c5779be74fa27955 on wip/alias-cell-ownership-20261030-185849; the hosted checks reported so far pass and no human review is recorded. The local continuation is on wip/issue2043-preflight-20261119-120058, with HEAD at ddd9e7f0b67762a51fd75f7a5cfca09bdf850688 and uncommitted runtime, regression, and design changes that are not on the PR. The checkout is not ready for UAT.

UAT and CPAN runs still need to clear the Moo accessor-weaken.t optree-reaping failure, Template Toolkit t/string.t counted-scalar replacement failure, and DBIx aggregate cleanup failures. Tracked Java regressions now cover destroyed hash/array outgoing edges, scalar-reference ownership, nested frames, owner-kind count consistency, stale aggregate owner slots, and reusable cleanup-probe state.

The latest unfiltered nice -n 19 timeout 3600 make test-unit passed in 7m43 (/private/tmp/issue2043-cleanup-probe-unit-extended.log); make check-links passed with 0 errors. Excel-Writer-XLSX/t/regression/chart_axis28.t passed 1/1 under the rebuilt candidate. Its diagnostics recorded 1,467 recursive-release roots across 360 worksets and 499 cleanup-target pre-scan roots across 125 reusable worksets. Owner-slot visits remained near 62,000. A four-second JFR allocation sample was noisy and did not establish a total allocation or throughput reduction. The full CPAN release-acceptance gate has not yet run on this candidate.

Allocation minimization plan

Allocation reduction is a completion requirement for sweep removal alongside lower scan cost and correct Perl lifetime behavior. Keep per-frame LIFO cleanup and the explicit owner ledger; use maintained O(1) owner-kind counts for common yes/no checks and retain identity indexes only where identities are needed. Reuse lazy per-frame queues, recursive-release worksets, and cleanup-target probe queues/visited sets across roots instead of allocating a collection per candidate element. The focused chart test's probe workset initializations fell from one per 499 calls to 125 reusable worksets, about 75% fewer for that path. The duplicate direct weak-slot pass was folded into the probe.

The next implementation step is to release the dying aggregate's outgoing owner slots directly, avoiding nested target pre-scans and repeated incoming-owner ledger scans while preserving shared-owner, weak-reference, and destructor-rescue behavior. Track owner-slot visits, pre-scan roots/worksets, queue growth, and append/drain counts separately. Use paired JFR and elapsed-workload measurements on the same CPAN target; the short current recordings are diagnostic evidence only, not a speedup claim. Keep changes only when allocation and scan costs both improve and the hash, array, scalar-reference, and nested-frame regressions pass.

Validation

Earlier merged UAT candidate
  • nice -n 19 timeout 1800 make passes on source commit 5c011dfa0e in 9m36 (/private/tmp/issue2043-merged-uat-make.log).
  • make check-links passes with 855 links checked and 0 errors (/private/tmp/issue2043-merged-uat-check-links-current.log).
  • The UAT launcher reports 5c011dfa0e; the smoke command prints UAT_OK.
  • dev/modules/dbix_failed_connect_releases_handle.t passes 3/3 and dev/modules/dbix_failed_connect_releases_storage.t passes 4/4 on JVM and interpreter; system Perl validation is recorded in the design handoff.
  • On the merged candidate, DBIx t/storage/error.t passes 49/49 with no leaks. t/storage/txn.t has one exception line-reporting mismatch at assertion 89, present before the ownership fix; it reports no leaked references, while system Perl passes 128/128. The interpreter CPAN attempts previously hit their 300-second limits.
  • The full eight-target make test-cpan-release-acceptance remains incomplete. The prior run reached DBIx failures and ended before Excel finished; the isolated Excel profiling run below was deliberately stopped after capturing JFR and is not a module pass.
  • Current hosted checks for PR head 75dd7d542f1f86bff930beb1c5779be74fa27955 pass; the PR is open and has no human review recorded.
Current local candidate
  • nice -n 19 timeout 3600 make test-unit passes in 7m43 after the reusable cleanup-target probe change (/private/tmp/issue2043-cleanup-probe-unit-extended.log).
  • The final documentation link check passes with 0 errors (/private/tmp/issue2043-check-links-final.log).
  • Focused Excel-Writer-XLSX/t/regression/chart_axis28.t passes 1/1; diagnostics and JFR are recorded at /private/tmp/issue2043-cleanup-probe-chart-axis28-metrics-29217.json and /private/tmp/issue2043-cleanup-probe-chart-axis28.jfr.
  • Full make and the eight-target make test-cpan-release-acceptance remain required on the final immutable candidate. Stop at an actionable CPAN failure, preserve its evidence, fix it, and restart the gate on a new candidate.

Remaining work

  1. Replace nested cleanup-target pre-scans and repeated incoming-owner checks with processing of the exact outgoing owner slots of dying arrays and hashes. Preserve weak observers, live shared aggregates, scalar-reference semantics, and destructor rescue.
  2. Measure scan visits, workset and queue allocations, sampled allocations, and elapsed time on the same Excel::Writer::XLSX workload; require paired evidence before claiming a broad speedup.
  3. Run focused regressions for the previously observed Moo, Template Toolkit, and DBIx failures, then nice -n 19 make and the complete eight-target CPAN release-acceptance gate on one immutable source candidate. Stop after an actionable failure, fix it before continuing, and rerun required gates on the resulting candidate. The fork failure remains waived by the user.
  4. Reconcile the tested local candidate with PR #1938, publish only after required gates pass, and wait for human review. Do not mark the checkout UAT ready or merge before final evidence.

After #2043 lands, resume #2037 archive-lifetime and #2034 localized-STDERR work separately.

主要語言
Perl
星號
67
分支
8
平均合併
5 小時 12 分鐘
30 天內合併 PR
131

環境準備

  • 提供 Dockerfile 或 Docker Compose 檔案
  • 沒有 Pull Request 範本
  • 閱讀貢獻指南

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。

fglock/PerlOnJava 的其他 Issue

查看 fglock/PerlOnJava 的全部 Issue

相似的 Issue

更多 Perl 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。