[antiguru]

9 of 14 in the interactive-runtime stack. Stacks on #38391, followed by #38393. Tracked by CPU-216. The design document is doc/developer/design/20261107_interactive_runtime.md in #40559.

Both runtimes reconcile each controller connection on their own. Without coordination, the interactive runtime can bind the peer of an index recreated at the same as_of to the slot of the incarnation the maintenance runtime has yet to drop, and then read that incarnation's frozen view.

• Worker::reconcile runs the compactions and drops it decides, then waits at a ReconcileBarrier it shares with the worker of the same local ordinal on the other runtime, then runs the creates. A waiting worker steps its timely worker but reads no command, so every worker of a runtime still executes the same commands in the same order. Solo skips the barrier.
• Connections are ordered by a ConnectionOrder that clusterd's multiplexing client records each Hello in before forwarding it. Global worker 0 of each runtime runs in process 0, looks up a connection's position there when it sequences the connection's commands, and the command channel delivers the position with them to every worker of its runtime. A worker never reads its own process's order, because the controller sends the other processes a connection's Hello but not its commands.
• A worker passes once the other side arrived at the same connection or a later one. A later one counts because the barrier keeps only each side's latest arrival, which the other side can overwrite before the waiting worker rechecks, and because a runtime's channel adapter drops a connection's queued commands once a response send to it fails, so one runtime can finish a connection the other never does.

Inert: enable_compute_interactive_runtime is off, so no replica runs two runtimes.

Tests:

• reconcile_barrier/tests.rs covers passing on the same connection, both reasons a later arrival counts, the wake of a parked waiter.
• server/tests.rs runs two runtimes' reconciliations: the interactive runtime creates only after maintenance dropped the old incarnation, and a worker whose process never recorded the connection still reconciles. sharing/tests.rs adds the same scenario at the registry.
• The clusterd-test-driver gains workflow_two_runtime_compute, which starts clusterd with both runtimes (Clusterd(interactive_compute=True)) and runs reconciliation.spec and the new two_runtime_recreate_index.spec with one and two workers. A peek's timeout-secs makes a read bound to a frozen incarnation fail instead of hang.

A reviewer should check the NOTE that the drop phase waits on no other worker, and the argument for later arrivals in the module doc of reconcile_barrier.rs.

Posted by Claude Code.

🤖 Generated with Claude Code