MESA Core · v0.9.3

Yeniden oluşturmanın tamamlanması

v0.9.3 kaynak: docs/adr/0014-v4-closure-and-staged-evolution.md

Bu rehberin Türkçe ve İngilizce sürümleri aynı konuyu açıklar. Bu sayfayla ilgili kaynak bölümleri en altta yer alır.

Bu kaynak, belirtilen tarihteki testleri kaydeder. Burada paylaşılması, testlerin bu site için yeniden çalıştırıldığı anlamına gelmez.

Yeniden oluşturmanın tamamlanması#

Bu belge, v0.9.3’in tasarım kararının gerekçesini açıklar. Sistemin bugün nasıl davrandığını güncel mimari ve çalışma belgeleri belirler. Artık geçerli olmayan kararlar, tasarımın nasıl değiştiğini anlamak için saklanır.

Gerekçeyi ve sonuçlarını okuyun#

Kararın durumu, gerekçesi ve sonuçları aşağıdaki özgün kaynakta yer alır. Genel çerçeve için karar dizinine ve mimariye bakın.

Kaynak belgedeki ilgili bölümler

Bu bölüm, kaynak belgenin bu sayfayla ilgili kısımlarını değiştirmeden gösterir. Kaynak içindeki bir bölüme giden bağlantıyı açtığınızda metin otomatik olarak görünür.

  • Status: Accepted for the rebuild boundary; the roadmap below is historical.
  • Partial supersession: ADR 0015 advances

bounded retrieval planning. The other future stages below are not implied

to be implemented. Current behavior is in architecture.

  • Baseline: MESA 0.7.x at c590188
  • Extends: ADR 0009, ADR 0012 and ADR 0013

Context#

The V4 canonical ledger and projection outbox exist, but a production-safe

projection rebuild also needs durable control state, one storage owner,

truthful capability reporting, deterministic verification and a reversible

cutover. LanceDB and Kùzu are currently shared by every dataset under a storage

root, so their physical layout cannot safely support tenant- or dataset-scoped

rebuilds.

Later architecture proposals depend on these closure guarantees. Implementing

them first would expand the number of storage and retrieval behaviours that

must be recovered without first establishing a reliable recovery boundary.

Decision#

The accepted rebuild boundary is independent of the historical feature order.

The following rules remain the operational constraints; current commands and

metadata behavior are maintained in the rebuild runbook.

The rebuild has the following hard boundaries:

  • It is an offline runner, not an API or worker background task.
  • MESA_V4_REBUILD_ENABLED defaults to false. Enabling submission does not

make an online rebuild safe; operators must drain work and stop the combined

runtime before running it.

  • Its only scope is the complete storage root. Requests for tenant- or

dataset-scoped rebuilds fail closed until vector and graph stores are

physically partitioned.

  • Canonical source/assertion content is not rebuilt. The runner rebuilds

LanceDB and Kùzu projections from canonical ownership records and validates

FTS integrity on a snapshot. Current operation/generation and representation

metadata updates are detailed in the rebuild runbook.

  • SQLite owns durable operation state and the single active projection

generation pointer. Cutover changes that pointer atomically. The previous

generation and the pre-run backup are retained until an explicit operator

cleanup outside this release.

  • Combined and rebuild runtimes acquire the same storage-root

writer lock before opening any writable store.

  • Provider/version mismatch, path or symlink ambiguity, insufficient disk,

parity failure and post-cutover health failure all fail closed. A failed

post-cutover check restores the retained generation pointer.

  • Dataset-bound MCP tools cannot submit a global rebuild. Administrative

operation APIs require a control-plane ADMIN principal and expose only

bounded, content-free status.

Closure gates#

Merge readiness requires deterministic V4 API contracts, migration and

repository contracts, rebuild fault injection, real-provider rehearsal where

available, formatting/lint/type checks and a content-free CI evidence

manifest. The manifest records schema and API identity, golden IDs, RRF

ablation, test results and small fixture-store counts without storing memory

content.

Production remains NO-GO after merge. A production GO additionally

requires a real-provider rehearsal on production-like disk and concurrency,

reviewed evidence and a 24-hour soak.

Consequences#

Operators get a resumable, auditable and reversible projection recovery path,

but the release deliberately does not offer online or scoped rebuilds. Storage

cost temporarily increases because backups and retained generations are not

automatically deleted. No online or scoped rebuild capability is implied.

Historical roadmap#

At baseline c590188, the proposed order was closure, focused DAO boundaries,

Ladybug, trusted assertion/bitemporal semantics, retrieval planning, MESA-Legal,

then community/global memory and hosted storage. This is a record of intent,

not an implementation checklist or a requirement to implement those stages.

ADR 0015 subsequently advanced bounded

query planning. Other stages require their own implementation and evidence;

current capabilities remain defined by the running API and architecture guide.

Rollback#

Disable new submissions, stop the offline runner, and keep mutation admission

closed while an operation is retryable. If a new generation was activated,

atomically restore the retained generation pointer before restarting the

combined runtime. Never delete the backup or previous generation as part of

automatic rollback.