v0.143.0

Two silent-loss fixes, both found by a blind audit and both measured before and after. A multi-schema Postgres sync could resume from a stopped-stream ALTER and apply rows against a target column that no longer matched the source — target reading 12:00:00 where the source read 21:00:00, at exit 0. And a Vitess/PlanetScale cold copy running with vstream_copy_table_parallelism >= 2 silently dropped statement-format DML instead of refusing it.

If you run multi-schema Postgres sync, or parallel VStream copy, take this one.

Fixed

A multi-schema Postgres sync now refuses a stopped-stream zone swap instead of priming (SLM-1d). The first-boundary session-zone-cast refusal compares an incoming schema against a seed — the shape the stream last committed under. The two single-stream reader-open sites installed that seed; the two multi-database sites wired the scope predicate and nothing else, and the fan-out never set a seed loader at all, so calling the helper there would have been a no-op. The measured harm, reproduced on real postgres:16 both before and after: stop the stream, ALTER TABLE ... TYPE timestamp under Asia/Tokyo, resume — the row lands, the target column is unchanged, and source and target disagree by the session offset with nothing reported. The fan-out now builds a namespace-partitioned seed on both the cold-start and warm-resume legs. Postgres is covered; a MySQL multi-database stream still cannot fire this refusal at all, which is now stated in the runbook's residual list rather than left out of it.

The concurrent VStream COPY pump refuses statement-format DML like its three siblings. dispatchCopyEvent had no statement-DML arm and fell through to a silent default, so under vstream_copy_table_parallelism >= 2 — the only condition where that pump is in use — statement-logged writes were ignored rather than refused. Under binlog_format=STATEMENT or MIXED the server emits no row events at all, which is precisely the loss the refusal exists to catch. The dispatcher universe is now derived from the source rather than hand-listed, and the environmental premise the whole class rests on — that the vendored vstreamer forwards these events rather than absorbing them — is pinned against the module's own source instead of living in a comment.

A table-filter pattern that matches nothing no longer cries wolf. v0.142.0 added TABLE-FILTER-PATTERN-UNMATCHED because an --exclude-table matching nothing fails open — the table you meant to keep out is copied at exit 0. Three false-fire arms have since been closed: the Postgres scope push-down (v0.142.1), engine-default _vt_* exclusions on PlanetScale, and multi-database runs, which called the filter door once per database and so reported a pattern naming a table in database B as dead while database A was being copied. A fan-out now stays quiet per pass and reports once, against the union of every selected database's tables.

Every remedy sluice prints can now be pasted. A refusal fires mid-incident and its remedy is the whole recovery path; sluice trigger setup has two required flags, so "re-run sluice trigger setup" produced a second refusal — "--dsn is required" — leaving you two failures deep on sluice's own advice. 51 messages across nine commands were fixed, and where the values are known the remedy now names your actual tables rather than a placeholder. A gate derives each command's required flags from the real CLI model, so this cannot regress quietly.

Compatibility

Drop-in from v0.142.1. No flag, format or schema change.

Remedy text changed in many error messages. If you match sluice's output with scripts or alerts, check anything keyed on the exact wording of a trigger setup / sync start / migrate remedy. Error codes and markers are unchanged.

One new refusal can fire where nothing fired before: a multi-schema Postgres sync resuming across a stopped-stream timestamptz/timestamp swap now refuses instead of priming. That is the fix. Apply the same ALTER on the target via the drained model and restart with the same --stream-id.

Who needs this

• Multi-schema Postgres sync — the silent-divergence fix.
• Vitess / PlanetScale with parallel VStream copy (vstream_copy_table_parallelism >= 2) — the silent row-drop fix.
• Anyone using --include-table / --exclude-table, especially on multi-database runs or PlanetScale.
• Anyone who has ever pasted a remedy out of a sluice refusal.

Development notes

This release was shaped by a full blind audit — five independent workers plus a reconciler — and the fixes above are its top findings. Two things are worth recording.

Two gates certified the defect they existed to prevent. One counted a comment as evidence, so disabling a warning while leaving its explanation in place kept the gate green. The other used a fixed lookahead window that ran past one call site into the next, so an unclassified site was satisfied by its neighbour's evidence. Both were mutation-proved by the audit and both now catch that exact mutant.

The release's own pre-tag review found a defect in the release's own new code, for the seventh time in eight releases — and this time the worst item was a remedy rewrite that made things worse: naming --dsn and --tables but not --source-driver meant a paste on the SQLite/D1 lane silently ran the Postgres installer, where before it had failed fast. Fixed at all seven sites before the tag.

Install

brew install sluicesync/tap/sluice
scoop bucket add sluicesync https://github.com/sluicesync/scoop-bucket && scoop install sluice
go install sluicesync.dev/sluice/cmd/sluice@v0.143.0
docker pull ghcr.io/sluicesync/sluice:0.143.0