Skip to content

Support for MergeAppend #736

@Fokko

Description

@Fokko

The so-called fast-appends are added in #349

It would be good to also consider adding merge-commits.

With the fast-append, a new manifest is written out and added to the manifest-list as mentioned in the spec. As the name suggests, this is the fastest way of appending new data, minimizing the chance of conflicts. Also, it works pretty well in the case of a commit, since only the manifest has to be rewritten in case of a conflict. The biggest drawback is that you create many manifests adding overhead in the long run (more calls to the object store than needed).

The merge-commit takes an existing manifest, adds the new entries to it, and replaces the old manifest in the manifest-list.

Having too few manifests is not good because it will lead to limited parallelization, but too many will add much overhead in terms of networking and parsing. The thresholds can be configured through configuration, and have some reasonable defaults:

image

The goal of this issue is to add MergeAppendAction next to FastAppendAction. This is not a trivial task since there are some caveats:

  • Each manifest is bound to a certain partition strategy, meaning that the partition-spec-id is stored in the Avro header, and they should be all the same.
  • When rewriting the existing manifests, the ADDED status must be changed to EXISTING, and the sequence numbers must be tracked correctly.

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions