0.1.13

0.1.13

2026-09-07

New

• ClickHouse source: Snapshot replication from ClickHouse.
• Scheduled snapshots: Snapshot only connectors can now repeat on a fixed interval or a timezone aware cron schedule.
• MCP server: Each agent exposes /mcp, so an MCP client can inspect and operate its connectors.
• Merge schedules: Send buffered CDC changes to the target on separate peak and off peak intervals. A long off peak interval lets a suspendable warehouse stop between merges and cuts its compute cost.
• Column hashing: Replace selected source column values with stable hashes, either the fast non cryptographic XXH3 128 or the keyed cryptographic HMAC SHA 256.

Core

• You can now enable history mode per table from the source catalog. The default is all tables.
• New SnapshotBatch task type resyncs a set of tables as one batch on both snapshot only and CDC connectors.
• Quoted identifiers that contain quotes, escapes, and other special characters are handled in more cases across sources and targets.
• Metrics:
  • New buffer.files.pending gauge reports about how many files wait in the buffer or are not yet processed by the target. The UI shows it as Pending files.
  • The metrics API uses bounded memory at startup.
• API: When you replace a connector, omitted secrets are kept.
• UI:
  • Refresh the catalog from the source catalog step.
  • The version indicator shows the build git sha.

ClickHouse

• Snapshot load modes:
  • Atomic swap publishes a snapshot through EXCHANGE TABLES.
  • Staged rename publishes a snapshot through ClickHouse RENAME TABLE. Use it when atomic swap is unavailable, for example without kernel support for renameat2.
  • Truncate clears the table before the load.
• Performance:
  • CDC upserts run against target tables in parallel.
  • Buffered CDC files are batched across flush windows into one load per table, up to 256 files at a time. Completed buffer files are deleted in bulk. Backlogs drain faster during catch up.

Apache Iceberg

Fixes a positional delete mode issue where external compaction during a snapshot restarted the snapshot.

BigQuery

• Supports placeholder mode for unchanged TOAST values from PostgreSQL.
• Rate limit, quota, and transient job errors are retried until BigQuery recovers.

Snowflake

Text values up to 128 MB and binary values up to 64 MB are supported.

Kafka

Broker storage errors (KafkaStorageError) are now retried.

PostgreSQL

• The replication slot's confirmed position now advances while the connector skips WAL for tables outside its catalog. Scoped down catalogs no longer hold WAL during catch up.
• Snapshot chunking falls back to physical page counts when row statistics are missing. Parallel chunking needs PostgreSQL 14 or later. Older versions snapshot each table as one chunk with a warning.
• Partitioned tables with leaf partitions in other schemas are now chunked.

SQL Server

Snapshot chunking now handles string primary keys (CHAR, VARCHAR, NCHAR, NVARCHAR).

Oracle

BFILE columns are supported in snapshots.

MongoDB

• Snapshot batches are now bounded by input bytes instead of row count alone. BSON documents can be far larger than a typical database row, so row count based batches used a lot of memory during schema inference. Many connectors can now snapshot at the same time on one agent. This applies to schema and schemaless modes.
• CDC now includes change stream pre-images for collections that have them enabled. Updates and deletes carry the state of the document before the change.