See:
https://emqx.atlassian.net/wiki/spaces/P/pages/612368639/open+e5.1+remove+auto+restart+interval+from+buffer+worker+resource+options
Current problem:
In 5.0.x, we have two timer options that control the state changing of buffer worker
resources: auto_restart_interval and health_check_interval.
- auto_restart_interval controls how often the resource attempts to transition from
disconnected to connected.
- health_check_interval controls how often the resource is checked and potentially moved
from connected to disconnected or connecting.
The existence of two independent timers for very similar purposes is confusing to users,
QA and even developers. Also, an intimately related configuration is request_timeout,
which can interact badly with auto_restart_interval if the latter is poorly configured:
requests may always expire if request_timeout < auto_restart_interval and if the resource
enters the disconnected state. For health_check_interval, we attempt to derive a sane
default that gives requests a chance to retry (if request timeout is finite, then the
resource retries requests with a period of min(health_check_interval, request_timeout /
3).
Another problem with the separate auto_restart_interval is that its default value (60 s)
is too high when compared to the default request timeout and health check, leading to the
problems described above if not tuned.
Proposed solution:
We propose to drop auto_restart_interval in favor of health_check_interval, which will be
used for both disconnected -> connected and connected -> {disconnected, connecting}
transition checks. With that, the resource will attempt to reconnect at the same interval
as the health check, which currently is 15 s.
Also, as two smaller changes to accompany this one:
- Increase the default request_timeout from 15 s to 45 s.
- Rename request_timeout to request_ttl.
Fixes https://emqx.atlassian.net/browse/EMQX-9905
Since calling `telemetry` is costly in a hot path, we instead collect
metrics inside the buffer workers state and periodically flush them,
rather than immediately as events happen.
Fixes https://emqx.atlassian.net/browse/EMQX-9367
For better user experience and performance for the average bridge, we
should change the default queue mode to `memory_only`, as was the
behavior of most bridges in e4.x. This leads to better performance
when message rate is high enough and the remote resource is not
keeping up with EMQX.
Also, we set the default segment size to equal max queue bytes.
Kafka Producer and Consumer bridges rely on this prefix for detecting
a dry run and avoid leaking atoms. At some point, this prefix was
changed, effectively disabling the check in Kafka Producer.
some telemetry events from wolff are discarded:
* dropped:
this is double counted in wolff,
we now only subscribe to the dropped_queue_full event
* retried_failed:
it has different meanings in wolff,
in wolff, it means it's the 2nd (or onward) produce attempt
in EMQX, it means it's eventually failed after some retries
* retried_success
since we are going to handle the success counters in callbac
this having this reported from wolff will only make things
harder to understand
* failed
wolff never fails (unelss drop which is a different counter)
With this, we avoid performing work or replying to callers that are no
longer waiting on a result.
Also introduces two new counters:
- `dropped.expired` :: happens when a request expires before being
sent downstream
- `late_reply` :: when a response is receive from downstream, but the
caller is no longer for a reply because the request has expired, and
the caller might even have retried it.
This makes the buffer/resource workers always use `replayq` for
queuing, along with collecting multiple requests in a single call.
This is done to avoid long message queues for the buffer workers and
rely on `replayq`'s capabilities of offloading to disk and detecting
overflow.
Also, this deprecates the `enable_batch` and `enable_queue` resource
creation options, as: i) queuing is now always enables; ii) batch_size
> 1 <=> batch_enabled. The corresponding metric
`dropped.queue_not_enabled` is dropped, along with `batching`. The
batching is too ephemeral, especially considering a default batch time
of 20 ms, and is not shown in the dashboard, so it was removed.
