spec-5.19 MG-B/MG-D: heap-ITL WAL delta v3 (close single-node write-tax gate)#11
Merged
Conversation
added 3 commits
June 29, 2026 20:51
…-logged median - M3: two-node peer-online single-writer write-tax measurement (real ClusterPair, strict quorum + shared_data; node0 writes while node1 is in quorum). REPORT ONLY: never asserts a threshold and never fails the single-node hard gate -- if the ClusterPair cannot boot/quorum/measure this run it passes with an explicit "unavailable" note. Addresses the 2-node write-path question without weakening the M1 single-node gate. - M1: emit the measured median write tax via diag() (reaches the CI log even on PASS; note() is swallowed by non-verbose prove) so the gate's headroom is visible without re-running the shard verbose. The HARD gate stays the single-node M1 tax <= 10% (rule 8.B).
…valid commit_scn, 48->40B/record) Closes the MG-B single-node write-tax blocker (rule 8.B). CI nightly measured the v2 (48B/record) tax at 10.62% > the 10% hard gate; an A/B that registered the delta 8B shorter passed the gate, confirming the 8B drop is sufficient. The write-time commit_scn (8B) is ALWAYS InvalidScn at every write-path ITL emit site: heap_insert / multi_insert / delete / lock / lock-chain / update old+new only ever stamp ITL_FLAG_ACTIVE / ITL_FLAG_LOCK_ONLY_ACTIVE transitions (the slot is not yet committed; COMMITTED stamping happens via the later commit-time / delayed-cleanout page mutation, FPI-logged, not via a write-path delta). Dropping an always-Invalid field is lossless. - heapam_xlog.h: new xl_heap_itl_delta_v3 (32B) + CLUSTER_ITL_DELTA_FORMAT_V3; keeps UBA (undo_segment_head moves 24->16), only commit_scn elided. StaticAsserts pin sizeof==32 and all field offsets. - cluster_itl.c: redo dispatches v1/v2/v3 by block format_version; v3 reconstructs commit_scn=InvalidScn. v1/v2 branches retained for backward WAL replay. consumed-bytes helper extended for v3. - heapam.c: all 8 write-path emit sites switch v2->v3 (drop the commit_scn assignment; register sizeof(v3)=32). - The COMMITTED-requires-valid-SCN redo guard still fires for any v3 delta that carries ITL_FLAG_COMMITTED -> fails closed (PANIC), so v3 can never silently install a committed slot with InvalidScn (8.A). - catversion 202606330 -> 202606340: fences an old binary from replaying v3-format WAL (unknown format_version -> redo PANIC). - Per mutating heap record: 8 + 40 == 48B -> 8 + 32 == 40B. - Tests: test_cluster_itl_wal v3 layout (T30-T33); D8 L6 invariant -> 40B (v2 40B retained for backward replay); t/329 MG-D model -> 40B + decision framing (v3 GO part shipped; same-block coalesce remains a follow-up).
…ative/two-node TPS + tax + unavailable reason) Output-visibility hardening for the MG-B report-only leg (the M1 single-node gate already diag()s its median): - M3 now diag()s the native single-node median tps, the two-node peer-online node0 median tps, and the two-node write tax % — captured in the CI log even on PASS / non-verbose prove, so the report-only numbers are never a silent black box. - When the 2-node measurement is unavailable, M3 prints the SPECIFIC reason (ClusterPair boot failed + error / peers not connected+in_quorum within timeout / pgbench init failed / no valid rounds / native baseline missing) instead of a generic "unavailable". - Still strictly REPORT ONLY: M3 never asserts a threshold and never fails the single-node hard gate. No behavior change to the M1 single-node ≤10% hard gate.
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
1 participant
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.
Closes the spec-5.19 MG-B single-node write-tax hard gate (rule 8.B). v2 (48B/record) measured 10.62% > 10% on CI nightly; an isolated A/B (register delta 8B shorter) passed the gate, confirming the always-Invalid 8B commit_scn drop is sufficient.
v3 = xl_heap_itl_delta_v3 (32B): keeps UBA, drops the always-Invalid write-time commit_scn. Redo dispatches v1/v2/v3 by block format_version (v1/v2 retained for backward replay; v3 reconstructs commit_scn=InvalidScn). All 8 write-path emit sites emit ACTIVE/LOCK only — verified no COMMITTED delta — so the drop is crash-recovery-lossless; the COMMITTED-requires-valid-SCN redo guard still fails closed. catversion bumped (fences old binary from v3 WAL).
Per mutating heap record: 8+40==48B -> 8+32==40B.
Verification: local build StaticAsserts pass (v3==32B), cluster_unit test_cluster_itl_wal 33/33 + stage5 acceptance 6/6. Full nightly dispatched on this branch for the real t/328 perf number + crash-recovery TAP (8.A).
🤖 Generated with Claude Code