ingestr v1.0.65 - Arrow, CDC, and OneLake Fixes

ingestr v1.0.65 was published on July 6, 2026, with a small set of changes that matter most around Arrow casting, CDC metadata, and OneLake path behavior. The main data path fix is in the data buffer: Arrow array offsets are now honored when casting, which matters when a slice is passed through a load path instead of a full array. It is not marked as a prerelease.

The full release notes and downloads are on the GitHub release page.

The most direct correctness change is databuffer: honor arrow array offsets when casting. The release notes also include a merge entry for an mmap Arrow slice offset bug, so the theme is clear enough without reading beyond this release: sliced Arrow arrays needed safer offset handling during casts.

For ETL jobs, this is not a cosmetic detail. Arrow arrays can represent views over shared buffers. If code treats a sliced array as though it starts at zero, a cast can read the wrong logical rows or mix values from outside the intended range. That kind of defect is painful because the shape of the batch can look valid while the content is shifted.

This release note does not claim a schema change or a new data type. It points to a narrower fix in the buffer layer. Operators who move Arrow backed batches through ingestr should treat v1.0.65 as a correctness patch, especially when upstream readers or intermediate transforms can pass slices rather than freshly materialized arrays.

The CDC item in v1.0.65 tags resume cursor lookup with ingestr_step. That gives a pipeline operator a better handle for tracing where cursor state is being read, especially when multiple steps touch the same destination metadata or when a job resumes after an interruption.

The note specifically says the annotation is for CDC resume cursor lookup. It does not say the cursor format changed, and it does not describe a migration. The practical effect is observability and attribution. When a CDC run has to explain why it resumed from a given point, step level tagging is a cleaner signal than a generic metadata lookup.

This also matters for BigQuery CDC users because the changelog includes a BigQuery CDC ingestr metadata merge entry. The notes do not give a broad BigQuery feature description, so it is better to read this as metadata plumbing around CDC rather than a new connector surface.

OneLake handling gets two entries. The project now keeps area directory materialization, and it skips managed lakehouse path prefixes. These are small words for a path sensitive storage integration. The important part is that ingestr is making clearer choices about which OneLake directories it should materialize and which managed prefixes it should leave alone.

That matters in lakehouse based ELT because a loader can create durable layout side effects. A wrong directory decision can make tables harder to discover, or worse, can place generated files in a path that the platform expects to control. The notes do not mention a breaking path migration. They do suggest operators should pay attention to OneLake destination behavior after upgrade if they rely on custom area layouts.

The release also adds a rule to parse table-string query params with the shared pkg/tablespec package. Shared table spec parsing is the kind of internal change that reduces edge cases across sources and destinations. It is not a new command by itself, but it can affect how table identifiers are normalized when users pass table settings through a URI style interface.

v1.0.65 adds an add-source agent skill for implementing new ingestr sources. That is a contributor workflow change, not an operator feature. It matters because source coverage is central to an ingestion tool, and a more explicit source authoring path can reduce drift between new connectors.

There is also a review cleanup around register.go registration and per source readParallel. The note is brief, and it reads like internal maintenance. Still, registration and per source read parallelism are not throwaway areas. They define how a connector is discovered and how it moves work through a run.

The careful read is that this release has a lot of infrastructure polish in the connector layer, but the release notes do not claim a new source. For platform teams, the concrete takeaway is that source addition, registration, and parallel read settings got attention in the same release as data buffer and storage path fixes.

• Release page: GitHub release page
• Repo: bruin-data/ingestr
• Tag: v1.0.65