Deletes from every store the phase writes to, the side being read first unless writeOrder is parallel.
Asks both stores what they hold for each id and returns the ids where
they disagree: a different version, or one side deleted and the other
not. Use it to find what a reported shadow failure left behind, then
write (or delete) those ids again from your source; the version rule
makes rewriting a copy a store already has harmless.
Pass ids in batches your stores can answer in one call. Both stores need
versions. The result is a snapshot: a write running at the same time
can show up as drift that is gone a moment later, so check again before
acting on a short list. After to-only the old store stops receiving
writes, so drift there is expected.
Returns the k nearest documents from the profile the phase reads,
passing filter to its store. In compare, a compareRate share of
queries also run against the other profile with the same filter, and
those calls wait for both.
Optionalfilter: FTakes effect for calls that start after it; calls already running keep the phase they started with.
Embeds each document with every profile the phase writes to, and upserts
the vectors. With the default writeOrder, the side being read goes
first and the other side is written only if that succeeds, so it never
holds a write the read side refused.
Sends writes and reads to two embedding profiles according to the current phase. One instance per process; the phase lives in memory, so every process in a fleet must be moved separately. The README's "Running more than one replica" section covers driving Migration.setPhase from a shared flag or config store.