Use rdkafka event API instead of the callback API

[scanterog]

This PR changes how rust-rdkafka interacts with librdkafka. Instead of using the callback API, it moves to the event API. This PR does not change any of the public rdkafka API, but it give us the chance to introduce some optimizations. Some of them might break the current API though and hence this PR does not go into that route yet. For example, consider #273.

Some other benefits:

• it allows to extend the lifetime of BorrowMessages. For example, consider the FutureProducer where we need to detach the borrowMessage on errors.
• it allows handling the rebalance events outside of the scope of the registered callback. This is specially useful for services using rust-rdkafka to build FFI libraries on top of it.

Performance wise, I've run kafka-benchmark I've got a pretty similar outcome for both the callback vs event based approach. I didn't find much differences here. I've also run long running services on our staging infra and we didn't find noticeable changes on cpu/rss on the producer/consumer side.

There's one main difference here though. Operation errors not related to consumption (like transient broker transport failures) are passed back to the user with the Event API. This is not the case for the Callback API. Details can be found confluentinc/librdkafka#4493. These errors should generally be considered informational as the underlying client will automatically try to recover from any errors encountered, the application does not need to take action on them. However, fatal errors are also propagated via the Event API. These indicates that the particular instance of the client (producer/consumer) becomes non-functional.

[mensfeld]

Hey, could you explain to a fellow maintainer (rdkafka-ruby here) what benefits it yields for the consumer? I clearly see benefits for admin and producer (where the delivery reports materialization is easier to handle) but I wonder about the consumer. I would appreciate some explanation as it has been a while since I started thinking about this but I have problems wrapping my head around the consumer part. Is it not possible to yield from rust in the context of rebalance callback? How will it operate on rebalanes where the revocation happens first and prior to it an action like offset storage could take place? Thanks 🙏

[scanterog]

Is it not possible to yield from rust in the context of rebalance callback? How will it operate on rebalanes where the revocation happens first and prior to it an action like offset storage could take place?

Hello @mensfeld! The rebalance handler does not really change here for rust-rdkafka. Is there anything in particular that gives that impression? Can you elaborate a little bit more about this?

[duarten]

LGTM in general!

[duarten]

We set op_timeout to std::cmp::min(timeout, min_poll_interval), so it's never going to be larger than timeout. Shouldn't we do something like:

let now = Instant::now();
let op_timeout = std::cmp::min(timeout, min_poll_interval);
// ...
let timeout = timeout.saturating_sub(now.elapsed());
if timeout == 0 {
    return None;
}

[scanterog]

It seems to work fine for the following common cases. Did I miss something?

• If the user provides a timeout set to 100ms, then the op_timeout will be 100ms (as min_poll_interval is 1sec by default). If no msgs or errors have been returned, then on this check op_timeout is equal to timeout and we return None.
• If the user provides a timeout set to 2 seconds (over the default for min_poll_interval), then op_timeout will be 1 second. Let's say we poll for 1 second and it does not return a msg/error, then op_timeout is less than timeout and now we do timeout -= op_timeout which set timeout to 1 second. If we get None messages/errors again, then op_timeout is equal to timeout and we return None.
• If the user provides timeout set to Never, then op_timeout is set to the default of 1 sec and we keep polling every second until we get a message or error.

[duarten]

I was concerned about something like the following:

• the user sets the timeout to 100ms.
• After 1ms, we get RD_KAFKA_EVENT_OFFSET_COMMIT event.
• We do timeout -= op_timeout, which puts timeout at 0.
• We exit.

[scanterog]

ah that's a good catch. Ok, I'll fix this.

[duarten]

Is this safe? Could we drop messages here and have the offsets committed to kafka?

[scanterog]

That's a good question. My understanding is that messages won't be forwarded here anymore. In this particular case we had to switch rd_kafka_consumer_close by rd_kafka_consumer_close_queue because rd_kafka_consumer_close will block until the consumer has revoked its assignment by calling Unassign or IncrementalUnassign on the rebalance handler. This translates on the consumer being blocked forever as there's no rebalance callback registered anymore.

Hence we had to move to rd_kafka_consumer_close_queue that allows us to asynchronously close the consumer. In this particular case, by calling Poll we can fwd the rebalance event to the rebalance method on the ConsumerContext following the API doc:

The application must poll/serve this queue until rd_kafka_consumer_closed() returns true.

[duarten]

Alright, sounds good to me.

[davidblewett]

Looks great!

[mensfeld]

@scanterog

Is it not possible to yield from rust in the context of rebalance callback? How will it operate on rebalanes where the revocation happens first and prior to it an action like offset storage could take place?

Hello @mensfeld! The rebalance handler does not really change here for rust-rdkafka. Is there anything in particular that gives that impression? Can you elaborate a little bit more about this?

This:

it allows handling the rebalance events outside of the scope of the registered callback. This is specially useful for services using rust-rdkafka to build FFI libraries on top of it.

Since rebalance callbacks are triggered from the thread that is polling the consumer queue (one or both with main), a closure in the callback should allow to run code "outside" (in a way). So I wondered how this is done via events API.

I am myself looking to move to the IO event API for producer and admin but not sure what benefits it yields for regular high level consumers.

[scanterog]

it allows handling the rebalance events outside of the scope of the registered callback. This is specially useful for services using rust-rdkafka to build FFI libraries on top of it.

Let me give some more context. Let's assume you provide some FFI layer on top of rust-rdkafka that allows you to create bindings for other langs (like Go/Python etc). Using the callback approach, we register a rebalance callback with rd_kafka_conf_set_rebalance_cb. This callback is responsible for running any *assign calls, such as rd_kafka_assign, within its scope. You cannot perform these *assign calls outside of this function. Consequently, it's challenging to have a rebalance callback in the wrapper language that can build up state based on rebalance events. While you can forward the assign/revoke events, you cannot call *assign, which means you cannot determine when to synchronize state with Kafka when partitions are assigned or revoked.

On the other hand, with the Event API approach, there are no such limitations. You receive the assign/revoke event and can delegate the *assign calls to an external caller. For rust-rdkafka, this translates to overriding the rebalance method of the ConsumerContext. This rebalance method can simply forward the event and delegate the *assign call to the wrapper language. We're not tied to the scope of the rebalance callback anymore.

[mensfeld]

@scanterog thank you for your explanation. That simplified things for me a lot 🙏