The integration plan

Public graph

Created about 9 hours ago, updated about 9 hours ago · Owners: Root

ID29553

What it is. Every merge proposal carries a plan: one step per node of your graph (the merge_steps table), saying what happens to it in the public graph. Prepare a merge writes the first version from the matches you choose.

The steps.

  • create: a new public node, with your text or a rewrite of it.
  • enrich: your text goes into an existing public node. The step holds that node's full new text, and a fingerprint of the text it was written against, so a change made since is noticed.
  • link: an extra parent edge between public nodes.
  • attach: your original graph hangs under the target as an attached graph.
  • skip: already covered; nothing happens.

Each step names its parents either as a public node or as another step of the plan, so new nodes can hang under each other and under existing ones at once.

Tried before. Proposing runs the whole plan inside a database transaction that always rolls back, so a plan that passes is exactly what approving will apply. Problems are shown on the step they concern: a parent that is gone or not public, steps whose parents refer to each other in a loop, text that changed underneath. From the same dry run, ChangeSet works out what the merge would do: new nodes, edited nodes with a diff, new connections. The AI first pass and the owners decide on that. The text is frozen once proposed.

Applied after. When every node it changes has an answer, the approved steps run for real, in one transaction: new nodes first (parents before children), then edits as new versions credited to you, then edges, then the attachment. If any step fails, nothing is applied and the proposal says why. Steps under nodes that were not approved are pruned first (see Approvals and pruning).

Undo. Each applied step can be rolled back on its own: see Roll back a merge.


Part of Merge: fork, edit, merge back and A proposal's life: all its parents are equal.