Skip to content

OpenADS mtfix75 full test kit - Windows x86 + Windows x64 + Linux x64 server + macOS universal

Pre-release
Pre-release

Choose a tag to compare

OpenADS mtfix75 cumulative test candidate

Includes the paused mtfix74 indexed-key Skip repair and fixture cleanup, plus append-record replacement batching and a matched Harbour RDD error bridge. Timing work is excluded. No live installation is changed.

Replacement boundary

APPEND BLANK remains its own acknowledged operation. On a successfully appended current record with valid generation and locks, the first replacement joins later replacements in one batch. Each explicit dbCommit remains a distinct boundary, including a key-only commit followed by a data commit. Movement, another append, unlock and close drain pending replacements before their next operation. Movement does not add a durable commit or release application locks. Clean commits and clean pending-field drains do not add a request. Existing-row first-field probing remains unchanged.

Own pending field reads use the buffered value before fetching a row. An old append acknowledgement without the required metadata keeps the prior first-field probe behavior. No new wire opcode is added. Server errors from missing columns and twin flushes are checked and returned; partial applied prefixes are not rolled back or automatically retried. A lost acknowledgement invalidates the connection rather than replaying an uncertain write.

Matched library requirement

The kit Harbour library must be rebuilt with patches/harbour/rddads-pending-fields.patch against Harbour4ec2e154ed92757418bc30af571ab90240ab3209. Relink Vouch with the updated librddads.a. A DLL swap alone does not add RDD error checks. New RDD code imports AdsFlushPendingFields; use the matching new DLL and import library. New RDD plus an old DLL is unsupported and must fail for the missing entry, not fall back to a forced commit. No Vouch source change is required.

The additive AdsFlushPendingFields(table) entry drains pending field assignments only: no commit, movement or unlock, and no request when clean. The RDD calls it before Goto's historical invalid-record compatibility path and before a new append, so prior write errors are not mistaken for navigation or ordinary append failures. Ordinary append errors retain the stock Harbour category/subcode/default flags. APPEND BLANK expands to dbAppend(); the command ignores the boolean. The default append-lock handler still sets NetErr() and allows dbAppend() to return false. Other errors follow the existing handler.

Evidence and limits

Local selected C++ and actual Harbour runtime tests pass. Harbour tests exercise successful batches, error boundaries, default append-lock behavior, ordinary append errors against stock baseline, deferred write failure before append, generic Goto5025 compatibility, refresh tails and SQL close callers. SQL close tests inject SQL-success/no-cursor before the actual RDD close path; they do not certify engine SQL execution with a simultaneous close fault. Flush-helper fault tests are synthetic, not a real disk-failure test. Windows and all-platform candidate gates must pass before publication. No universal any-error, all-platform parity, rollback, safe-retry, Vouch-runtime or production-readiness claim.

The x86 full default test gate retains the existing197 failing cases and exact assertion signatures, with no new failures or harness errors. It is not a clean full-suite pass. Linux and macOS retain their configured full ctest gates; Windows x64 retains binary/package checks rather than a native full unit-suite claim. All prior auth, refresh, commit, bottom, diagnostics and mixed-server gates remain. Existing73 releases are frozen. Remote key-Skip repair requires the new server; a new DLL cannot repair an old server's hidden cursor.

OpenADS mtfix75 full cumulative test kit

Pritpal Bedi. Fork bedipritpal/OpenADS from FiveTechSoft/OpenADS.

Frozen73 plus a bounded engine repair for indexed-key edit followed by relative Skip. Successful GoCold/index maintenance marks a lazy per-index re-anchor; Skip re-finds the current physical record under its new key before walking forward/backward. Parked tags retain their own pending state; replacement index objects do not inherit it. Duplicate runs have no new arbitrary cutoff. Failed re-anchor invalidates the cursor and returns the error. Server Skip now checks cursor, ordered-twin and natural-table failures instead of issuing a normal ACK.

AAA to AAZ then Skip(1) must land on BBB, following Harbour DBFCDX GoCold/current-key refresh ordering. Local repair requires the75 DLL; remote repair requires the75 server. Old73 client/new75 server is included in the compatibility gate. New75 client/old72 or73 server retains the old server defect; a client DLL cannot repair that hidden engine cursor. No wire-format change, automatic unlock or forced hard flush. User locks remain held; later explicit commit still handles pending durability. No Vouch/new C edit, count-caller, read-ahead or OPEN-policy repair. Frozen releases remain unchanged.

No extra re-anchor seek is added to indexed append/commit itself. The cost is deferred to relative navigation after relevant key maintenance. Tests check hot and explicit-commit edits, forward/backward, key crossing, duplicate runs above1024, Skip0, scopes, local descending, multiple tags, conditional state lifetime, disk bytes, retained locks and injected failures. Remote descending API remains unsupported and is not certified. Conditional membership navigation is not broadly redesigned; lifecycle bookkeeping and physical Goto outside a FOR tag are checked. Existing partial data/index maintenance on failure is not rolled back; an error does not certify a safe retry. GoTop/Seek masked dirty failures and other handlers remain outside this repair. This is not full Harbour parity or production-readiness certification.

Retained71 commit fusion,72 GoBottom metadata and73 caller diagnostics remain cumulative. Diagnostics default OFF. The x86 gate requires1821 executed cases, exactly the existing197 baseline failing names/assertion signatures, all12 new indexed-key Skip cases passing, and zero harness errors. Linux/macOS retain their full configured ctest gates. Windows x64 retains PE/package checks, not a native unit-suite claim. New12-case local matrix passes7707 assertions; combined targeted local runner passes106 cases/25841 assertions. These local checks do not replace all-platform release gates.

Retained72 behavior and limits

Frozen71 plus capability-gated same-landing scoped key-count metadata for an ordinary ordered GoBottom whose internal key-position dependency is cold. One bottom reply instead of bottom plus GetKeyCount in that case. No Vouch speed claim. Existing matched71 commit fusion remains cumulative.

A 16-byte magic/status/count/table-id prefix wraps the otherwise unchanged row/boundary/pair reply only when requested by a capable new client. Successful count is consumed once to complete this GoBottom's last-key-position calculation. It does not seed persistent count caches. Natural order, warm count, existing fused-order and certified pair paths retain their existing behavior. Old peers retain B0 fallback.

No new commit, lock, unlock, first-field, RDD GoCold, on-disk format, Vouch source, new C, read-ahead or OPEN policy change. Ordered count uses the same ABI active order/scope/SET DELETED semantics as the prior internal B0 dependency. Failed optional count is not certified; legacy count path remains. Malformed/missing/wrong-table metadata fails this read without replaying navigation. Existing ignored navigation/fused-count return policies outside this new certification are not repaired or certified.

Explicit count APIs are not rewritten. Baseline71 already has persistent RemoteIndex key-count caches and optional physical-count caching unless OPENADS_FRESH_COUNTS=1. This candidate is not blanket freshness certification. Because the single-use GoBottom completion does not populate those caches, a later scoped count can now require its own request; it is not silently served from this new metadata. Existing empty-scope EOF/BOF distinctions remain and are tested separately, not normalized.

The owner's rule that record movement should return full table state is recorded as a design direction. This candidate supplies only the missing GoBottom count dependency. It does not implement a broad all-movement protocol overhaul or certify all table state/freshness. Production readiness and other open issues remain separate.

Net trips depend on the COMPLETE call sequence: bottom alone2->1; bottom followed by later explicit same-index count2->2(request shifts); bottom plus later explicit table-handle count3->2. The owner's captured final bottom has no later explicit count. No universal net-saving or elapsed-latency claim.

Inherited explicit refresh re-reads the ordered ABI twin used to pack the response and
propagates twin refresh errors. No Skip-freshness B, refill optimization, Vouch
source rewrite, on-disk format change or browse-speed claim.
Added userspace reads measured in a 2200-row/31-byte fixture:65542DBF+1028CDX
bytes per full-block refresh;2674DBF+1028CDX final partial. Not physical disk or
elapsed production latency. Duplicate reads retained for correctness/error safety.
Shared pre-existing relation-child re-seek and stored-descending bottom/EOF gaps
remain outside this repair. No full relation/descending semantics claim.
Source CI including Windows actual Harbour RDDADS/independent DBFCDX writer:
https://github.com/bedipritpal/OpenADS/actions/runs/41083765147
Windows actual Harbour marker REFRESH-PEER-HARBOUR-PASS verified there.
Not proof of Vouch compiled RDDADS/ForceStable chain.
Windows x86 is not a clean unit-suite pass: same197 baseline failing cases with
identical recorded assertion results/sourcefiles required.1821 cases must execute,
zero harness errors; new indexed explicit-refresh regression must pass113 assertions,
prior Tim2e regression must still pass304 assertions, and the retained mtfix68 append-metadata suite must pass181 assertions in 6 cases.
All prior cumulative fixes/toolchains and auth/HRB/diagnostics gates retained.
Main/upstream and running installations untouched. No automatic deployment.
Client diagnostics default OFF. Explicit opt-in remains required.
2a customer-confirmed on reported setup;2b current fix awaits customer confirmation;
2c C8 repro and2d round-trip work remain open. No universal flashing claim.
macOS uses Intelmac15/Xcode16.4, universal slices, minOS14.0 metadata and full tests.
ARM slice and macOS14 runtime not independently tested; targetmetadata is not proof.
Windowsx64 has PE/package checks, not a nativeGCC unit-suite pass.

All four platform kits together; SHA256 sidecars/manifests and source pins.
Windowsx86 Qt4.9.2/GCC10.2 and ownerbinutils2.35.1 unchanged.
Public mock HRB only, not realcustomer auth. Test-only, no automatic deployment.
Use matched client/server for the mtfix64 repair layer. mtfix68 added append metadata; mtfix69/70 were client-only diagnostics. mtfix71 commit fusion retained; mtfix72 GoBottom server capability retained; mtfix73 caller diagnostics retained; mtfix75 adds engine indexed-key Skip re-anchor and checked errors. Remote repair requires server75; old servers retain their prior behavior.

Final verification

All five build/test jobs in https://github.com/bedipritpal/OpenADS/actions/runs/42583559761 passed. Automatic publication alone failed due workflow-token permissions; these exact unchanged green-run artifacts are published here. Source pin: 4429a98. Workflow pin: 21c6d1c. Actual Windows and Linux Harbour runtime and Windows import/missing-export checks passed. Frozen70/71/72/73 compatibility matrices passed. Windows x86 remains the same197 baseline failures, not a clean full-suite pass. Some inherited packaged notes refer to the previously paused74 repair layer; this cumulative75 contains that layer and is the matched kit to use. No74 kit was published.