D-WAL: LanceCycleWriter vs MemWAL measured; capped session cache, version clock, copy-free temporal projections by AdaWorldAPI · Pull Request #1562 · AdaWorldAPI/lance-graph

plan: wal-shootout-v1 — pre-register LanceCycleWriter vs lancedb 0.40 MemWAL (D-WAL-1..3)

claude

committed

beb9add

D-WAL-1: validation probe (V1-V4) for LanceCycleWriter vs the lancedb 0.40 MemWAL path; D-WAL-2/3 harness

claude

committed

d08dbde

D-WAL-1 results: both paths crash-atomic and lose nothing acknowledged; MemWAL writes invisible to other readers until flush or replay

claude

committed

a47cb94

D-WAL-2 results: per-cycle cost of LanceCycleWriter vs MemWAL at three cycle sizes

claude

committed

d6dc46c

D-WAL: rustfmt the probe, harness and example

claude

committed

b012485

WAL shootout: cross-check D-WAL-1/2 against the lance 13 MemWAL spec

claude

committed

de99538

D-WAL-3 results (truncated soak); board: D-WAL entry, #1555 arc and state

claude

committed

257baa0

D-WAL-4 probe: metadata-cache growth per Lance version (default vs capped session)

claude

committed

a22bde3

D-WAL-4: L's RSS growth is Lance's per-version manifest cache (V^2/2), measured

claude

committed

c1708c2

D-WAL-5: LanceCycleWriter opens and writes on a session with an 8 MiB metadata cache

claude

committed

a658b36

D-WAL-6: VersionClock — a TemporalPov resolves to one row range by two array reads

claude

committed

7ea51a3

D-WAL-5: cover the reopen path in the session-cap test

claude

committed

fe13cd7

D-WAL-6: version clock cost probe (v4: 4.2 ns per rows_at, scan 127 us at 400k rows)

claude

committed

fa8ab52

D-WAL-7: index forms of deinterlace and local trajectories; copying forms become wrappers

claude

committed

0c34302

wal_shootout: box the L writer (clippy large_enum_variant after the session field)

claude

committed

c076de8

board: D-WAL-5/6/7 done, D-LNC-7 decided (L default, MemWAL optional), VersionClock in the contract inventory

claude

committed

0b27755

plan: cite cycle_sink's Dataset::write sites by function, not line number (citation-decay)

claude

committed

bd3c377

CI: gate wal_shootout_probe with cfg(feature) instead of required-features; temporal index forms refuse > u32::MAX rows

claude

committed

628d51a

D-WAL probe: V1 must bracket the commit boundary, V2 must reach the target ACK

claude

committed

e38779f

Navigation Menu

Repository navigation

D-WAL: LanceCycleWriter vs MemWAL measured; capped session cache, version clock, copy-free temporal projections - #1562

D-WAL: LanceCycleWriter vs MemWAL measured; capped session cache, version clock, copy-free temporal projections#1562
AdaWorldAPI merged 19 commits into
mainfrom
ccr-9f3b8075-74u47l-wal-shootout

Commits on Oct 15, 2026

plan: wal-shootout-v1 — pre-register LanceCycleWriter vs lancedb 0.40 MemWAL (D-WAL-1..3)

D-WAL-1: validation probe (V1-V4) for LanceCycleWriter vs the lancedb 0.40 MemWAL path; D-WAL-2/3 harness

D-WAL-1 results: both paths crash-atomic and lose nothing acknowledged; MemWAL writes invisible to other readers until flush or replay

D-WAL-2 results: per-cycle cost of LanceCycleWriter vs MemWAL at three cycle sizes

D-WAL: rustfmt the probe, harness and example

WAL shootout: cross-check D-WAL-1/2 against the lance 13 MemWAL spec

D-WAL-3 results (truncated soak); board: D-WAL entry, #1555 arc and state

D-WAL-4 probe: metadata-cache growth per Lance version (default vs capped session)

D-WAL-4: L's RSS growth is Lance's per-version manifest cache (V^2/2), measured

D-WAL-5: LanceCycleWriter opens and writes on a session with an 8 MiB metadata cache

D-WAL-6: VersionClock — a TemporalPov resolves to one row range by two array reads

D-WAL-5: cover the reopen path in the session-cap test

D-WAL-6: version clock cost probe (v4: 4.2 ns per rows_at, scan 127 us at 400k rows)

D-WAL-7: index forms of deinterlace and local trajectories; copying forms become wrappers

wal_shootout: box the L writer (clippy large_enum_variant after the session field)

board: D-WAL-5/6/7 done, D-LNC-7 decided (L default, MemWAL optional), VersionClock in the contract inventory

plan: cite cycle_sink's Dataset::write sites by function, not line number (citation-decay)

CI: gate wal_shootout_probe with cfg(feature) instead of required-features; temporal index forms refuse > u32::MAX rows

D-WAL probe: V1 must bracket the commit boundary, V2 must reach the target ACK