What shipped for launch
First of these. The format will stay the same: what changed, why it matters, and what's still missing. A changelog that only lists wins is marketing with a version number on it.
Bring files in without writing a loader¶
Data Tools → Tools → Ingest Files. Drop a CSV, TSV, Parquet, JSON or NDJSON file (gzipped or zstd-compressed if you like, up to 50 MB) and land it in your warehouse as a new table, or append it to one that already exists.
The schema is inferred from the file and shown to you before anything is written, so you can correct a column type rather than discover it afterwards.
It's aimed at reference data: lookup tables, mappings, the spreadsheet export someone emails you every quarter. It's deliberately not a bulk loader. For large or recurring loads a pipeline reading from object storage is the right tool, and the 50 MB ceiling exists to make that obvious now rather than at 400 MB.
Monitoring, split into four questions¶
Monitoring used to be one list. It's now four reports, because "what happened" turned out to be four different questions:
- Runs — step-level history with the real failure context, including the database's own error message
- Workflow Executions — whole-workflow runs and what triggered each one: scheduled, API, or ad-hoc
- Audit Trail — every agent tool invocation, joined to its plan node, task and session
- Transfer Usage — pipeline data stored plus files ingested, against your plan's allowance
The Dashboard leads with Steps w/ Errors and opens on it. Failures are what you want on landing; everything else is one click away.
The workflow editor got its edges back¶
The workflow builder now matches the SQL editor: same tab shape, same toolbar, an environment selector beside Run instead of a dialog in the way.
The canvas now sizes itself to the window, which fixes something embarrassing. The zoom and lock controls sit in the canvas's bottom-left corner, and at a fixed height they could end up below the fold on a laptop. They were there the whole time. You just couldn't reach them.
Connectors say what they need¶
Every connector in Add Integration now carries a one-line description that states what credential it wants. That sounds cosmetic. It fixes a specific confusion: several connectors are the same database reached a different way, and nothing told them apart.
PostgreSQL — Host / Port — configured field by field. Connect with host, port, database, username and password.
PostgreSQL (connection string) — the same databases, reached through a single URI instead of separate fields.
The search box reads those descriptions too, so typing "connection string" finds the connectors that take one. The docs gained a Getting your credentials section linking to each vendor's own page, including the one that catches everyone: Google Cloud Storage needs an HMAC key, not a service- account JSON.
Not here yet¶
- Pipeline tests. Assertions that gate a deployment to production. The card is in the product marked Coming Soon rather than hidden, because it's the most- asked-for thing and we'd rather show the gap than pretend the roadmap is a secret. Until it lands, environment separation is the safety half of testing and there's no correctness half.
- Agent steps. Calling a model as part of a pipeline run, the way a SQL step runs a query. The connectors exist; the step type doesn't. This is separate from the Analyst, which is available now and needs no key from you.
- Alerting. Failure notifications have delivery history coming; channels are configured in Organization settings for now.
Coming next¶
Nested pipelines, meaning a workflow used as a step inside another workflow, is the next substantial piece. Today it's rejected at execution time rather than silently mis-run. That's the correct failure, but not a satisfying one.
🧩 Keywords¶
release notes, file ingestion, monitoring, workflows, duckdb, product update