This change is part of the following stack:

• chore: deploy patch #4324
  • chore: flatten runtime config #4326
    • fix(api): make actor destroy wait for ack #4327
      • fix(udb): add txn tagging #4328
        • chore: bump rivetkit protocol version #4329
          • fix(ups): formalize subjects #4334
            • fix(api): properly dedup name list responses #4337 ◀
              • fix(pb): clear actor reschedule ts on alloc #4338

_{Change managed by git-spice.}

🚅 Deployed to the rivet-pr-4337 environment in rivet-frontend

Service	Status	Web	Updated (UTC)
frontend-inspector	😴 Sleeping (View Logs)	Web	Mar 3, 2026 at 12:19 am
website	😴 Sleeping (View Logs)	Web	Mar 3, 2026 at 12:17 am
frontend-cloud	❌ Build Failed (View Logs)	Web	Mar 3, 2026 at 12:10 am
mcp-hub	✅ Success (View Logs)	Web	Mar 3, 2026 at 12:09 am
ladle	❌ Build Failed (View Logs)	Web	Mar 3, 2026 at 12:07 am

Code Review

Summary

The use of IndexMap/IndexSet for deduplication across datacenter fanouts is a good approach and correctly solves the core dedup problem. However, there are a few issues worth flagging.

Bug: Limit applied before sort in list_names.rs and runners.rs

engine/packages/api-public/src/actors/list_names.rs (lines 63-69)
engine/packages/api-public/src/runners.rs (lines 101-108)

In both files, .take(limit) is applied to the iterator before sorting. IndexMap iterates in insertion order (the order results arrived from the datacenter fanout). .take(limit) selects the first limit entries in that arbitrary insertion order, then sorts them. The result is a sorted subset of a non-deterministic slice — not the first limit items alphabetically.

For example, if DCs return names ["z", "a", "b", "c", "d"] and limit=3, you get ["a", "b", "z"] instead of ["a", "b", "c"]. This also breaks cursor-based pagination since the cursor will point to the wrong position in the full ordered set.

The pre-existing code had the correct order: sort then truncate. The fix should collect all deduplicated names, sort, then apply the limit. Same issue applies in runners.rs.

Potential regression: cursor filtering removed in list.rs actor_ids path

engine/packages/api-public/src/actors/list.rs (the actor_ids branch)

The manual cursor filtering was removed. And fetch_actors_by_ids always constructs its peer query with cursor: None hardcoded (utils.rs:90), so the cursor is never forwarded to the peer. Clients using actor_ids + cursor for pagination will silently receive incorrect results (the same first page on every call).

If cursor pagination is intentionally unsupported for the actor_ids path, returning an explicit error when both actor_ids and cursor are provided would be safer than silently ignoring it.

Minor: redundant actors.truncate(limit) in list.rs

After passing Some(limit) to fetch_actors_by_ids, the result is still truncated to limit again (line 122). fetch_actors_by_ids sorts but does not truncate, and with multiple DCs each returning up to limit results the aggregated count can exceed limit. The truncate is correct and necessary — just worth a clarifying comment since it looks redundant at first glance.

pnpm add https://pkg.pr.new/rivet-dev/rivet/@rivetkit/cloudflare-workers@4337

pnpm add https://pkg.pr.new/rivet-dev/rivet/@rivetkit/framework-base@4337

pnpm add https://pkg.pr.new/rivet-dev/rivet/@rivetkit/next-js@4337

pnpm add https://pkg.pr.new/rivet-dev/rivet/@rivetkit/react@4337

pnpm add https://pkg.pr.new/rivet-dev/rivet/rivetkit@4337

pnpm add https://pkg.pr.new/rivet-dev/rivet/@rivetkit/sql-loader@4337

pnpm add https://pkg.pr.new/rivet-dev/rivet/@rivetkit/sqlite-vfs@4337

pnpm add https://pkg.pr.new/rivet-dev/rivet/@rivetkit/traces@4337

pnpm add https://pkg.pr.new/rivet-dev/rivet/@rivetkit/workflow-engine@4337

pnpm add https://pkg.pr.new/rivet-dev/rivet/@rivetkit/virtual-websocket@4337

pnpm add https://pkg.pr.new/rivet-dev/rivet/@rivetkit/engine-runner@4337

pnpm add https://pkg.pr.new/rivet-dev/rivet/@rivetkit/engine-runner-protocol@4337

commit: d88911e