Zero 1.6

Postgres 17 Replication Failover and Performance

Installation

Copynpm install @rocicorp/zero@1.6

Features

• Postgres 17 Logical Replication Failover: zero-cache now creates replication slots with an ordinal naming scheme (e.g. zero_0_a, zero_0_b) so they can be registered with synchronized_standby_slots for logical replication failover. TODO: add docs link.
• Litestream Region: Added ZERO_LITESTREAM_REGION for deployments in non-standard AWS partitions like GovCloud (thanks @ericykim!).
• Optional args in custom queries/mutators: Custom query and mutator execution functions now treat args as optional when the args type already allows undefined (thanks @0xcadams!).

Performance

• Faster EXISTS subqueries via the new Cap operator, which lets SQLite skip ORDER BY for non-flipped EXISTS children
• Bulk-insertion optimization in Replicache via putMany, speeding up large sync patches (3-5x faster for typical sync batches, up to 53x for construction)
• Batch deletes and upserts in SQLiteStore writes (~7-9x faster on 1000-put commits)
• Parallelize I/O during pull and rebase
• Heap-based k-way merge in fetchMergeSort (O(log K) per row vs O(K)), with a new mergeSortedStreams utility
• Initial sync progress reporting uses pg_class estimates instead of full table scans
• De-dupe SQLite requests in flip-join when children want the same parent

Fixes

• Returning to an app after stale-tab GC or CVR purge caused a full page reload; now the Zero instance rotates in place
• "Row already exists" errors after an IVM advance failure could mask the original error and continue through corrupt branch state, also fixed for IVMBranch.fork()
• "Row already exists" assertion failures during poke processing caused by putMany rebalancing duplicating entries across adjacent BTree children
• Initial sync could fail or take hours on large databases because progress reporting did full COUNT(*) and SUM(pg_column_size(...)) scans
• Deadlock between post-initial-sync changeLog reset and a live replication-manager during non-disruptive resync
• Zombie ViewSyncers could accumulate in the active-client-groups metric when clients disconnected before initConnection resolved
• ConcurrentModificationException is now classified as a Rehome so the client reconnects instead of erroring
• zero-cache startup errors during change-streamer init were not published to subscribers
• TypeError: Expected string at context.query. Got null when handling DDL events with NULL current_query()
• Repeated initial-sync failures could exhaust the replication-slot name pool; cleanup now runs preemptively under the management lock, and inactive slots are deleted together with their replicas row so a stuck slot doesn't keep claiming a name
• Replication slot creation timeouts crashed the server during backfill retries; backfill timeouts now only error after the maximum retry backoff is reached
• Shadow sync threw when a synced table could not be queried by ZQL; it now silently ignores the table to match prod behavior
• WebSocket errors are now logged as warnings instead of errors, since they reflect client or upstream issues rather than server faults
• Inspector now caches AST and metrics for deleted queries so they remain accessible after eviction

Breaking Changes

• Inspector per-query hydration metrics format changed: Per-query hydration metrics (query-hydration-server-ms) are now reported as a plain number (most-recent hydration time in ms) instead of a TDigest histogram, and the per-query metrics type was renamed from ServerMetrics to QueryServerMetrics. If you have custom tooling reading inspector metrics, you'll need to update it. The protocol version was bumped from 50 to 51 to reflect this; MIN_SERVER_SUPPORTED_SYNC_PROTOCOL remains at 30, so 1.6 servers remain compatible with older clients.