feat(facade): Basic RESP support in ioloopv2, part 2

[dranikpg]

Second part of #6870, testing on CI

Implement very basic admin message processing for ioloopv2

[dranikpg]

Some tests timed out

[dranikpg]

Not that are tests have good coverage, but those results are not bad

FAILED 😰  dragonfly/connection_test.py::test_reply_count[df_factory0] - assert 3 <= 2
FAILED 😰  dragonfly/connection_test.py::test_timeout[df_factory0] - AssertionError: assert 2 <= 1
 +  where 2 = len([{'addr': '127.0.0.1:33574', 'age': '52', 'db': '0', 'fd': '15', ...}, {'addr': '127.0.0.1:33560', 'age': '52', 'db': '0', 'fd': '14', ...}])
FAILED 😰  dragonfly/connection_test.py::test_pipeline_cache_only_async_squashed_dispatches[df_factory0] - assert 2520 == 0
FAILED 😰  dragonfly/connection_test.py::test_client_migrate_no_conn_leak[df_factory0] - TimeoutError: Timed out!
FAILED 😰  dragonfly/memory_test.py::test_throttle_on_commands_squashing_replies_bytes[df_factory0] - assert 0 > 0
 +  where 0 = len([])
==== 5 failed, 496 passed, 50 skipped, 75 deselected in 1409.74s (0:23:29) =====

[dranikpg]

I disabled the v2 loop for redis and kept everything as is, rounding it up here

[augmentcode]

🤖 Augment PR Summary

Summary: Extends the experimental IoLoopV2 path toward basic RESP/control-path support.

Changes:

• Prevents AsyncFiber-based control-path processing when ioloop_v2_ is enabled and instead wakes the main loop via io_event_.
• Updates IoLoopV2() to wake on pending dispatch_q_ items and drain them inline, flushing replies and notifying pubsub backpressure waiters.
• Adds !ioloop_v2_ guards to migration/async-fiber related paths to avoid unsupported interactions.
• Minor parsing-path adjustments to keep V2 compatible with the revised control-path handling.

_{🤖 Was this summary useful? React with 👍 or 👎}

[augmentcode]

Review completed. 1 suggestion posted.

Comment augment review to trigger a new review at any time.

[augmentcode]

Since ioloop_v2_ no longer launches async_fb_, the early return for cc_->conn_closing can drop CheckpointMessages even after SendCheckpoint() increments the BlockingCounter, potentially causing waits/timeouts. Consider ensuring checkpoints are either always decremented on closing v2 connections or SendCheckpoint only increments the counter if the message is guaranteed to be queued/processed.

Severity: high

_{🤖 Was this useful? React with 👍 or 👎, or 🚀 if it prevented an incident/outage.}

[Copilot]

Pull request overview

Updates facade::Connection to better integrate the experimental IoLoopV2 path by avoiding the legacy AsyncFiber wake/launch logic and instead waking IoLoopV2 directly and draining control-path (dispatch_q_) messages inside IoLoopV2.

Changes:

• Gate AsyncFiber-related behaviors (launch/wake/migration handling) when ioloop_v2_ is enabled.
• Add dispatch_q_ wake condition to IoLoopV2 and drain/flush control messages from within IoLoopV2.
• Disable RequestAsyncMigration() when ioloop_v2_ is enabled.

---

You can also share your feedback on Copilot code review. Take the survey.

[glevkovich]

How about setting here the async_dispatch flag (with a cleanup)?

if (!dispatch_q_.empty()) {
  cc_->async_dispatch = true;
  absl::Cleanup reset = [this] { cc_->async_dispatch = false; };

  while (!dispatch_q_.empty()) {
    auto msg = std::move(dispatch_q_.front());
    dispatch_q_.pop_front();
    std::visit(AsyncOperations{reply_builder_.get(), this}, msg.handle);
    UpdateDispatchStats(msg, false /* subtract */);
  }
}

[glevkovich]

Here we jump back to the beginning of the while loop, giving the dispatch queue strict priority over command parsing and execution. This may cause starvation - especially with a high volume of pubsub messages, the parsed command queue might never be served.

At a minimum, I suggest adding a TODO comment here to acknowledge the issue for a future task, or fixing it on the spot as part of this PR.