v0.7

Released October 16, 2026 · Preview v0.7.0

v0.7 adds persistent pipeline drafts, bounded data preview, node-level execution settings, managed Cloud sessions, and a source-only Db2 preview. It also improves PostgreSQL log release, restart recovery, and connector integrity checks. Use this preview with non-production systems and sample data. The server runs as a single member by default; cluster mode remains opt-in.

See the GitHub releases for native CLI archives, individual .sha256 files, checksums.txt, and platform-minimums.txt.

Before upgrading

• Back up persistent state and its keyring together. Stored Source configurations are encrypted in both deployment modes. The original keyring is required to read them. The store now has migrations through version 12; an older server that cannot read that version refuses startup.
• Review execution settings. settings.parallelism and settings.batch_size remain parseable but have no execution effect. Use the applicable node's execution settings instead.
• Retain the Cloud working volume for rollback. The Cloud image now runs as root and can create root-owned files. An earlier UID-10001 Cloud image cannot safely reuse that volume unchanged. The standard self-hosted image retains UID 10001. Both image profiles use Ubuntu 24.04.

Follow Upgrade to v0.7 for backup, verification, and rollback steps.

What changed

Web authoring and preview

• Save incomplete pipeline drafts. Save edits before they form a runnable pipeline, then compile and publish when ready. Saving, publishing, and starting are separate operations. Publishing does not restart an existing run. See Build a pipeline.
• Preview sample results before publishing. Data Preview compiles candidate resources, reads bounded source samples, and runs an isolated pipeline. It does not write to the destination. Results can use MongoDB Extended JSON or logical JSON, depending on the target codec. This differs from draft preview, which only compiles the configuration.
• Distinguish intent from observation. Status separates the last runtime observation from the requested state and UI display state. Saved drafts and known pipelines without an observation no longer require a fabricated runtime state. See Observe a pipeline.

Execution and recovery

• Configure execution per node. Set a cluster-wide target parallelism and batch policy where execution is accepted. Requested and effective widths can differ; source readers currently remain single-processor. See Pipeline execution settings.
• Release PostgreSQL WAL from durable progress. Shared capture confirms changes after they enter the durable shared log; direct capture waits for target confirmation. Clearing the last dependent pipeline attempts to release its replication slot. Snapshot-only reads avoid creating a CDC slot. See PostgreSQL.
• Recover from confirmed progress. Restart fixes cover slower sinks, PostgreSQL heartbeat boundaries, shared-capture takeover, joins, and pending restarts during startup. Missing replay history produces a recovery error rather than silently skipping changes. See Capture modes and recovery.

Cloud and connectors

• Use managed Cloud sessions. Cloud deployments verify exchanged identities and Cluster-owned sessions. Self-hosted deployments retain local authentication. Source configurations are encrypted in both modes. See Authentication.
• Evaluate Db2 as a source. Read a Db2 LUW snapshot and ongoing changes. Native CDC requires Linux x86_64 and the documented runtime preparation. Db2 targets are refused in both deployment modes, including after an upgrade. See Db2.
• Keep native connectors usable across operations. Connection tests, discovery, and pipelines reuse an artifact's class loader, avoiding repeated native-library loading failures. Updated Oracle and Db2 download metadata matches the published preview artifacts.

Other fixes

• MongoDB full loads honor on_full_load: clear and fail while preserving collection indexes and options. Recovery does not repeat a completed clear.
• Connection tests and schema discovery wait for the server to finish instead of reporting a client read timeout after 30 seconds. Cancellation and other request budgets remain separate.
• Unreadable connector uploads return connector.artifact-unreadable with HTTP 400 and the artifact name. Altered cached or stored artifacts are refused before loading.
• Stopping backpressured or shared captures releases their writers, cursors, and late snapshot buffers without disrupting the remaining consumers.
• The managed views connection cannot be deleted while stored pipelines depend on it. Self-hosted SRV metadata connections preserve the effective authentication database when deriving the views connection.
• Connected CLI diagnostics report both CLI and server versions. Quiet pipelines with settled parallel writes no longer report a false frontier gap.
• Scheduled GitHub Actions installs skip the default install-telemetry request; an explicit telemetry endpoint remains authoritative.

Known limits and issues

This release remains an evaluation preview.

• Delivery is at-least-once. Snapshot and CDC overlap can duplicate rows on insert-only targets. A separate snapshot-to-stream positioning window can miss changes the connector never reads; at-least-once delivery does not cover that window.
• Cluster traffic has no authentication or TLS. Keep the member port on a private, protected network.
• Join accepts a restricted SQL shape. Use supported equality joins with unique dimension keys. Duplicate dimension keys replace earlier matches; there is no uniqueness guard or 1:N fan-out. Keyed writes require all fact primary-key columns as direct selected columns. See Join for the full limits.
• Older stores can refuse CDC recovery. A store written by the published 0.4.5 image can retain legacy consumer identities and return capture.recovery-progress-unproven. Preserve the store and backup; do not clear recovery progress as a workaround. See #740.
• Shared PostgreSQL capture can fail on restart. A retained-state restart can return connector.capture-failed with a slot-occupied error. Preserve state and inspect the slot owner before retrying. See Replication slots and recovery.
• Concurrent registration can create ambiguity. Register connectors through one entry member. This avoids concurrent seeding but does not repair an already duplicated registry. See #687.
• Shared inbound metrics can include another pipeline's tables. Do not use those counts alone to assess delivery. See #719.
• Live Atlas acceptance remains deferred. Four external Atlas acceptance classes are deferred pending a managed test environment. MongoDB-based checks do not establish live Atlas acceptance. See #702.
• Db2 remains a preview. This release has no automated live-Db2 acceptance lane. Validate your exact connector artifact and database before use.

Get started

• Run the Web Quickstart, then build a pipeline with your own data.
• Read Upgrade to v0.7 before changing an existing server.
• Install the CLI for scripts and AI integrations.
• Use the v0.7.0 release assets for binaries and checksums.
• Review the complete changelog for the commit and PR history.