Reactor jumping to the wrong Scheduler?

I've got a flux I'm processing in steps using flatMaps and calling blocking code with a Mono.fromCallable(). When looking in the run logs, the retires are running in the parallel pool for some reason. Why is that, and how can I get them to execute in the scheduler that the group is running on? This is basically the workflow below, starting with a stream of messages, group by some key, then for each group, I want to run them in their own Scheduler with a max concurrency of 6 (in this case). What am I doing wrong?

messageStream
    .groupBy(this::grouper)
    .flatMap(group ->
        Scheduler scheduler = Schedulers.newBoundedElastic(6, 1, "ChildPool:"+group.key());
        return group
              .publishOn(scheduler)
              .flatMap(t1 -> Mono.fromCallable(() -> chevronOne(t1))
                   .retryBackoff(MAX_RETRIES, Duration.ofSeconds(1), Duration.ofSeconds(60))
                    .onErrorResume(e -> doErrorChevronOne(m, e, t1))
              )
              .flatMap(t2 -> Mono.fromCallable(() -> chevronTwo(t2))
                   .retryBackoff(MAX_RETRIES, Duration.ofSeconds(1), Duration.ofSeconds(60))
                   .onErrorResume(e -> doErrorChevronOne(m, e, t2))
               )

Here are a couple of lines from the logs showing the Scheduler jumping:

2020-08-12 15:33:20.289 INFO [ChildPool:1004-21] e.u.t.t.f.l.transfers.ITestTransfers:582
2020-08-12 15:33:20.889 INFO [parallel-2] e.u.t.t.f.l.s.TransfersService:475 - ***** DOING chevronTwo ****

publishOn only guarantees that given Scheduler will be used by the following operator. retryBackoff schedules the retry on the parallel scheduler by default, hence the result.

You must always be explicit about the scheduling if you need to handle something on a specific Scheduler.